EdgeFirst/visdrone-yolo11-det

YOLO11 Detection on VisDrone — EdgeFirst Model Zoo

YOLO11 detection models trained on VisDrone2019-DET (10 classes: pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, motor) and validated on its val split (548 images) on real edge hardware with the EdgeFirst Profiler, which runs the pipeline and computes the accuracy on the device itself. Each row in the tables below cites the EdgeFirst Studio validation session (v-XXXX) that produced the measurement.

VisDrone is drone footage: most objects are a few pixels across, so small-object AP (APs) leads every table, followed by the full COCO-style metric set. Every model is published in four input variants that trade time per image for small-object accuracy. For the same family trained on COCO, see EdgeFirst/yolo11-det.

Part of the EdgeFirst Model Zoo. New here? Introducing the EdgeFirst Model Zoo covers what is in the zoo, why every figure links to its validation session, and how to reproduce any of it on your own hardware.

Training experiment: View on EdgeFirst Studio — dataset, training configuration, metrics, and exported artifacts.

---

Input variants

Each model and size ships in four variants, one folder each. They differ in the input the model sees and in the weights it was trained with, and each one buys small-object AP with time per image:

• 640 is the baseline: one 640×640 pass over the whole frame.
• 1280 shows what a bigger input alone buys: one 1280×1280 pass over the whole frame.
• 1344x768 gives the best result: weights trained on 640×640 tiles run natively on the whole frame at 1344×768, where memory allows.
• tiled (SAHI) slices the frame into 640×640 tiles, runs every tile, and merges the detections. It serves constrained targets, and images larger than VisDrone's ~8 MP, while keeping the benefit of tiled training.

Variant	Folder	Input	Weights	Inference
640	640/	640×640	Full-frame	One pass over the whole frame
1280	1280/	1280×1280	Full-frame	One pass over the whole frame
1344x768	1344x768/	1344×768	Tiled	One pass over the whole frame
tiled	tiled/	640×640 tiles	Tiled	Sliced inference (SAHI): every tile, then merge

Time per image

Every latency and FPS figure on this card is per image, whatever the image strategy: one 640² pass, one 1280² pass, one 1344×768 pass, or all the 640² tiles of an image plus slicing and merging. The four variants therefore compare directly.

For the tiled variant, time per image is roughly the number of tiles times the time per tile when tiles run one at a time, plus the cost of slicing the frame and merging the detections. Running tiles concurrently helps when the accelerator has spare capacity or when CPU pre- and post-processing can overlap NPU inference; it does not help when one tile already saturates a single NPU, or when memory is the limit (for example the NPU memory pool on the Toradex Verdin i.MX 95). Tiles overlap so that an object cut at a tile border is detected whole in the neighbouring tile; the overlap adds tiles, and therefore time, and produces duplicates that the merge step removes.

---

Reference accuracy — ONNX FP32

Accuracy ceiling for each variant and size, measured against VisDrone2019-DET val (548 images) with COCO-style metrics. Quantized and compiled artifacts are graded against the reference of the same variant.

Variant	Size	Params	APs	AP	AP50	AP75	APm	APl	AR@100	Source
640	Nano	2.6M	9.99%	18.46%	32.19%	18.08%	27.82%	44.26%	30.99%	v-1973
640	Small	9.4M	13.28%	22.23%	38.04%	22.30%	32.74%	45.25%	35.14%	v-1977
640	Medium	20.1M	17.02%	26.38%	43.90%	26.87%	38.30%	51.16%	39.64%	v-1980
1280	Nano	2.6M	19.75%	28.44%	46.70%	28.73%	39.68%	49.07%	43.78%	v-1974
1280	Small	9.4M	23.91%	32.86%	52.89%	34.20%	44.45%	54.52%	48.05%	v-1979
1280	Medium	20.1M	27.73%	36.42%	57.89%	38.04%	48.25%	55.54%	51.44%	v-1981
1344x768	Nano	2.6M	20.70%	29.95%	48.59%	30.68%	41.50%	50.94%	45.03%	v-1976
1344x768	Small	9.4M	25.45%	34.62%	55.38%	35.62%	46.66%	51.67%	49.86%	v-197e
1344x768	Medium	20.1M	29.36%	37.81%	59.76%	39.80%	49.88%	56.59%	53.03%	v-1988
tiled	Nano	2.6M	19.46%	28.10%	47.32%	28.15%	39.23%	46.33%	43.62%	v-1975
tiled	Small	9.4M	23.43%	32.35%	53.28%	33.13%	44.08%	47.11%	47.48%	v-197b
tiled	Medium	20.1M	26.74%	35.08%	56.84%	36.47%	46.78%	52.16%	50.12%	v-1985

---

On-target validation results

Each row is one EdgeFirst Studio validation session. Click the Source link to inspect the full session — model artifact, dataset version, parameters, per-stage Perfetto trace, and the host hardware description.

Row conventions in the tables below:

• Rows whose Δ APs cell reads ref are the ONNX FP32 reference runs each quantized or compiled measurement is graded against.
• Δ APs vs FP32 is the small-object AP change against the same variant and size's reference, in percentage points.
• Rows whose Δ APs cell carries a ⚠ measure more than 10 percentage points of APs below the same variant's float reference: below our expectations for that platform. The numbers are real measurements, reproducible from the linked Studio session, and we are investigating them.
• Not supported: … entries, in the short table under each variant's results, are configurations this edition could not produce or run, with the reason: a converter, memory or runtime limitation of that target, not a missing measurement.
• Runtime is the inference stack that executed the model; Decode is what the exported graph emits (Smart lifts the operations that quantize worst out of the graph and needs a decoder implementing that cut, the EdgeFirst HAL in EdgeFirst deployments; Logical keeps the head's outputs separate and named; — means no quantized decode split). Full converter documentation: EdgeFirst model conversion.
• Platform-label suffixes. (EVK) / (FRDM) / (Phytec) / (Verdin) / (Ezurio) name the development board a session ran on. — latency / — throughput mark the two pipeline configurations of the NXP i.MX 95 Neutron and NXP Ara240 targets: serial inference for the lowest latency, or several inference workers for the highest FPS.
• Time per image (ms) is the end-to-end latency for one image, from the moment the frame is available to the moment its results are ready — preprocess, inference (every tile, for the tiled variant), postprocess and merge. Image acquisition (file load and JPEG decode) and writing results to disk are excluded. Realized FPS is the measured steady-state rate of finished images over the full validation pipeline.
• The full metric set (AP, AP50, AP75, APs, APm, APl and AR at 1, 10 and 100 detections and per object size) for every row is under Full metrics below each table.

640 — 640×640, full-frame weights

Size	Platform	Runtime	Decode	APs	Δ APs vs FP32 (pp)	AP	AP50	Time per image (ms)	Realized FPS	Source
Nano	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	9.99%	ref	18.46%	32.19%	9.67	307.0	v-1973
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	9.86%	-0.13	18.37%	32.13%	8.63	336.2	v-19cc
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	3.42%	-6.57	10.41%	21.14%	314.14	4.6	v-1c33
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	3.42%	-6.57	10.41%	21.14%	101.54	10.7	v-1b35
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	3.42%	-6.57	10.41%	21.14%	101.60	10.7	v-1b36
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	8.32%	-1.67	16.38%	29.53%	110.81	11.4	v-1b3a
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	9.32%	-0.67	17.84%	31.49%	61.53	38.1	v-1761
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	9.32%	-0.67	17.84%	31.49%	80.11	47.1	v-1a0e
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	8.78%	-1.21	16.79%	30.09%	46.17	40.4	v-1ae7
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	3.49%	-6.50	10.66%	21.53%	52.06	22.6	v-1ae1
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	3.51%	-6.48	10.68%	21.57%	53.14	22.0	v-1adf
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	8.78%	-1.21	16.79%	30.08%	72.34	62.2	v-1ae8
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	3.49%	-6.50	10.66%	21.53%	80.77	57.3	v-1ae4
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	3.51%	-6.48	10.68%	21.57%	79.81	57.5	v-1ae0
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	8.78%	-1.21	16.79%	30.08%	70.64	64.2	v-188c
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	9.32%	-0.67	17.84%	31.49%	65.30	36.0	v-176f
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	9.32%	-0.67	17.84%	31.49%	85.15	42.5	v-1a18
Nano	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	5.52%	-4.47	13.78%	25.94%	17.05	107.4	v-1676
Nano	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	5.48%	-4.51	13.78%	25.94%	18.31	194.0	v-1928
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	8.80%	-1.19	17.17%	30.55%	24.05	51.4	v-162c
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	10.29%	+0.30	19.01%	32.96%	35.32	176.2	v-1910
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	8.14%	-1.85	16.66%	30.19%	34.12	183.8	v-1912
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	3.92%	-6.07	11.45%	22.92%	9.97	495.3	v-160e
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	9.29%	-0.70	17.78%	31.45%	12.15	272.3	v-1610
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	9.85%	-0.14	18.20%	31.90%	20.73	223.4	v-1612
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	8.98%	-1.01	17.28%	30.70%	11.68	295.1	v-1d3b
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	3.60%	-6.39	11.04%	22.10%	11.93	477.8	v-1d3c
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	9.84%	-0.15	18.24%	31.91%	17.50	235.5	v-1d3d
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	9.08%	-0.91	17.53%	31.39%	15.17	214.9	v-1d3f
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	9.85%	-0.14	18.25%	31.94%	14.74	213.6	v-1d3e
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	13.28%	ref	22.23%	38.04%	11.66	244.1	v-1977
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	13.18%	-0.10	22.13%	37.93%	9.07	273.8	v-19d1
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	4.21%	-9.07	11.81%	24.04%	412.99	4.3	v-1c4f
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	4.21%	-9.07	11.81%	24.04%	181.55	5.7	v-1b7a
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	4.21%	-9.07	11.81%	24.04%	181.31	5.7	v-1b7e
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	10.93%	-2.35	19.25%	34.09%	190.86	5.9	v-1b81
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	12.19%	-1.09	20.67%	36.55%	83.63	20.5	v-177c
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	12.19%	-1.09	20.67%	36.55%	154.22	33.0	v-1a1e
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	11.55%	-1.73	19.99%	35.41%	72.94	19.7	v-1b19
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	4.51%	-8.77	12.36%	25.08%	81.83	13.4	v-1b16
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	4.51%	-8.77	12.37%	25.08%	82.54	13.2	v-1b13
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	11.55%	-1.73	19.99%	35.42%	138.37	33.3	v-1b1a
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	4.51%	-8.77	12.37%	25.08%	132.49	32.1	v-1b18
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	4.51%	-8.77	12.37%	25.08%	128.69	33.3	v-1b15
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	11.55%	-1.73	19.99%	35.41%	139.52	33.3	v-18a3
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	12.19%	-1.09	20.67%	36.55%	86.80	20.4	v-1795
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	12.19%	-1.09	20.67%	36.55%	163.62	32.0	v-1a27
Small	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	7.47%	-5.81	16.37%	30.80%	21.02	83.5	v-1604
Small	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	7.46%	-5.82	16.37%	30.80%	18.83	143.1	v-194c
Small	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	12.32%	-0.96	21.18%	37.07%	45.01	25.0	v-163c
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	13.57%	+0.29	22.45%	38.57%	42.04	129.6	v-15cb
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	12.11%	-1.17	20.36%	36.30%	35.98	175.6	v-15cc
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	5.49%	-7.79	13.45%	27.05%	9.01	381.0	v-1688
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	12.34%	-0.94	21.17%	36.84%	12.98	308.3	v-168b
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	13.22%	-0.06	22.15%	37.98%	18.90	168.0	v-168c
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	11.86%	-1.42	20.62%	36.33%	12.98	313.2	v-1d79
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	4.56%	-8.72	12.50%	25.43%	9.57	378.5	v-1d7a
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	13.00%	-0.28	21.91%	37.69%	20.18	167.7	v-1d7b
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	12.34%	-0.94	21.46%	37.58%	17.75	157.0	v-1d7d
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	13.04%	-0.24	21.91%	37.74%	16.83	155.3	v-1d7c
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	17.02%	ref	26.38%	43.90%	21.37	151.6	v-1980
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	16.88%	-0.14	26.26%	43.86%	13.41	187.7	v-19d5
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	4.28%	-12.74	12.36%	24.95%	595.26	3.0	v-1c71
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	4.28%	-12.74	12.36%	24.95%	334.44	3.0	v-1bf0
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	4.28%	-12.74	12.36%	24.95%	334.64	3.0	v-1bf6
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	10.52%	-6.50	20.27%	36.67%	343.17	3.1	v-1bfd
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	11.31%	-5.71	21.05%	38.46%	148.92	9.4	v-17a1
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	11.31%	-5.71	21.05%	38.46%	382.14	11.7	v-1a2f
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	10.78%	-6.24	20.52%	37.80%	130.14	9.0	v-1b5c
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	4.41%	-12.61	12.59%	25.37%	141.88	7.4	v-1b58
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	4.44%	-12.58	12.61%	25.42%	144.47	7.2	v-1b52
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	10.78%	-6.24	20.52%	37.80%	360.06	11.8	v-1b60
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	4.41%	-12.61	12.59%	25.38%	347.86	11.7	v-1b5b
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	4.44%	-12.58	12.61%	25.42%	345.69	11.8	v-1b56
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	10.78%	-6.24	20.52%	37.80%	360.02	11.8	v-18ad
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	11.31%	-5.71	21.05%	38.46%	149.99	9.5	v-17c8
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	11.31%	-5.71	21.05%	38.46%	379.09	11.8	v-1a32
Medium	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	6.34%	-10.68 ⚠	15.57%	30.09%	35.88	39.0	v-1629
Medium	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	6.34%	-10.68 ⚠	15.57%	30.09%	32.18	46.6	v-1952
Medium	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	11.63%	-5.39	21.33%	38.71%	80.74	13.2	v-1643
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	17.33%	+0.31	26.75%	44.54%	65.77	70.8	v-193e
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	10.30%	-6.72	19.86%	36.79%	48.90	104.4	v-193f
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	5.86%	-11.16	14.49%	28.81%	12.26	213.1	v-16b2
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	12.85%	-4.17	22.72%	40.45%	16.67	223.0	v-16b3
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	17.04%	+0.02	26.38%	43.91%	32.41	94.2	v-16b5
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	11.18%	-5.84	20.79%	38.10%	18.22	230.3	v-1db2
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	4.92%	-12.10	12.68%	25.85%	15.18	216.9	v-1db3
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	16.64%	-0.38	26.06%	43.57%	31.99	92.0	v-1db4
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	16.16%	-0.86	25.68%	43.51%	27.78	90.5	v-1db6
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	16.67%	-0.35	26.07%	43.68%	27.41	87.5	v-1db5

Full metrics — 640

Size	Platform	Runtime	Decode	Precision	APs	AP	AP50	AP75	APm	APl	AR@1	AR@10	AR@100	ARs	ARm	ARl
Nano	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	9.99%	18.46%	32.19%	18.08%	27.82%	44.26%	8.62%	23.13%	30.99%	21.28%	44.32%	60.08%
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	9.86%	18.37%	32.13%	17.90%	27.68%	44.33%	8.59%	23.04%	30.89%	21.17%	44.16%	60.14%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	3.42%	10.41%	21.14%	9.44%	16.72%	32.13%	6.05%	15.63%	21.91%	12.35%	35.02%	46.02%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	3.42%	10.41%	21.14%	9.44%	16.72%	32.12%	6.05%	15.63%	21.91%	12.35%	35.02%	46.02%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	3.42%	10.41%	21.14%	9.44%	16.72%	32.12%	6.05%	15.63%	21.91%	12.35%	35.02%	46.02%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	8.32%	16.38%	29.53%	16.06%	24.93%	34.97%	7.83%	21.31%	28.58%	19.09%	41.55%	50.66%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	9.32%	17.84%	31.49%	17.34%	27.11%	42.66%	8.40%	22.84%	30.37%	20.56%	44.06%	59.06%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	9.32%	17.84%	31.49%	17.34%	27.11%	42.66%	8.40%	22.84%	30.37%	20.56%	44.06%	59.06%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	8.78%	16.79%	30.09%	16.21%	25.64%	39.68%	8.13%	21.84%	29.38%	20.07%	42.61%	54.82%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	3.49%	10.66%	21.53%	9.57%	17.40%	36.59%	6.30%	16.14%	22.41%	12.73%	36.39%	49.63%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	3.51%	10.68%	21.57%	9.61%	17.40%	36.63%	6.32%	16.18%	22.45%	12.72%	36.45%	49.63%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	8.78%	16.79%	30.08%	16.21%	25.64%	39.68%	8.13%	21.84%	29.38%	20.07%	42.61%	54.82%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	3.49%	10.66%	21.53%	9.57%	17.40%	36.62%	6.30%	16.14%	22.41%	12.73%	36.39%	49.63%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	3.51%	10.68%	21.57%	9.61%	17.40%	36.63%	6.32%	16.18%	22.45%	12.72%	36.45%	49.63%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	8.78%	16.79%	30.08%	16.21%	25.64%	39.69%	8.13%	21.84%	29.38%	20.07%	42.61%	54.82%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	9.32%	17.84%	31.49%	17.34%	27.11%	42.66%	8.40%	22.84%	30.37%	20.56%	44.06%	59.06%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	9.32%	17.84%	31.49%	17.34%	27.11%	42.66%	8.40%	22.84%	30.37%	20.56%	44.06%	59.06%
Nano	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	5.52%	13.78%	25.94%	13.34%	22.02%	39.52%	7.19%	18.84%	25.70%	16.57%	38.43%	50.02%
Nano	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	5.48%	13.78%	25.94%	13.35%	22.02%	39.37%	7.19%	18.84%	25.70%	16.57%	38.43%	50.02%
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	8.80%	17.17%	30.55%	16.80%	26.31%	44.23%	8.22%	21.78%	29.27%	19.61%	42.46%	56.26%
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	10.29%	19.01%	32.96%	18.79%	28.64%	46.87%	8.69%	23.86%	31.65%	21.62%	45.47%	61.30%
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	8.14%	16.66%	30.19%	16.19%	25.81%	40.73%	8.05%	21.80%	29.34%	19.47%	42.90%	57.51%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	3.92%	11.45%	22.92%	10.35%	18.90%	40.38%	6.55%	16.62%	23.02%	13.12%	36.94%	54.22%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	9.29%	17.78%	31.45%	17.17%	27.15%	43.31%	8.37%	22.41%	30.05%	20.44%	43.33%	60.23%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	9.85%	18.20%	31.90%	17.79%	27.49%	42.19%	8.55%	22.97%	30.70%	21.02%	43.91%	57.46%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	8.98%	17.28%	30.70%	16.65%	26.30%	41.90%	8.21%	22.21%	29.78%	20.37%	42.55%	57.52%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	3.60%	11.04%	22.10%	9.76%	18.22%	37.11%	6.31%	16.06%	22.39%	12.70%	35.79%	50.03%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	9.84%	18.24%	31.91%	17.77%	27.52%	43.34%	8.61%	22.94%	30.74%	21.12%	43.83%	58.26%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	9.08%	17.53%	31.39%	16.81%	26.85%	41.78%	8.30%	22.02%	29.55%	20.13%	42.09%	56.59%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	9.85%	18.25%	31.94%	17.74%	27.48%	43.39%	8.60%	22.95%	30.76%	21.14%	43.85%	57.80%
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	13.28%	22.23%	38.04%	22.30%	32.74%	45.25%	9.68%	26.63%	35.14%	25.95%	48.03%	58.89%
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	13.18%	22.13%	37.93%	22.24%	32.67%	45.19%	9.66%	26.46%	34.94%	25.72%	47.88%	59.12%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	4.21%	11.81%	24.04%	10.69%	19.38%	34.78%	6.53%	17.60%	23.91%	13.81%	38.19%	48.14%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	4.21%	11.81%	24.04%	10.69%	19.38%	34.78%	6.53%	17.60%	23.91%	13.81%	38.19%	48.14%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	4.21%	11.81%	24.04%	10.69%	19.38%	34.78%	6.53%	17.60%	23.91%	13.81%	38.19%	48.14%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	10.93%	19.25%	34.09%	19.21%	28.61%	38.58%	8.72%	24.47%	32.58%	23.36%	45.35%	54.07%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	12.19%	20.67%	36.55%	20.38%	30.46%	48.02%	9.03%	25.70%	33.99%	24.75%	47.00%	64.70%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	12.19%	20.67%	36.55%	20.38%	30.46%	48.02%	9.03%	25.70%	33.99%	24.75%	47.00%	64.70%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	11.55%	19.99%	35.41%	19.62%	29.53%	41.83%	8.89%	24.64%	32.94%	24.03%	45.41%	58.79%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	4.51%	12.36%	25.08%	10.96%	20.05%	40.00%	6.74%	17.98%	24.66%	14.75%	38.76%	53.99%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	4.51%	12.37%	25.08%	10.97%	20.07%	40.03%	6.74%	17.99%	24.66%	14.74%	38.77%	54.03%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	11.55%	19.99%	35.42%	19.62%	29.53%	41.83%	8.89%	24.64%	32.94%	24.03%	45.41%	58.79%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	4.51%	12.37%	25.08%	10.96%	20.05%	40.00%	6.74%	17.98%	24.66%	14.75%	38.76%	53.99%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	4.51%	12.37%	25.08%	10.97%	20.07%	40.03%	6.74%	17.99%	24.66%	14.74%	38.77%	54.03%
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	11.55%	19.99%	35.41%	19.62%	29.53%	41.83%	8.89%	24.64%	32.94%	24.03%	45.41%	58.79%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	12.19%	20.67%	36.55%	20.38%	30.46%	48.02%	9.03%	25.70%	33.99%	24.75%	47.00%	64.70%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	12.19%	20.67%	36.55%	20.38%	30.46%	48.02%	9.03%	25.70%	33.99%	24.75%	47.00%	64.70%
Small	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	7.47%	16.37%	30.80%	15.68%	25.72%	39.96%	8.08%	21.93%	29.66%	20.25%	42.72%	53.50%
Small	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	7.46%	16.37%	30.80%	15.68%	25.72%	39.97%	8.08%	21.93%	29.66%	20.25%	42.72%	53.50%
Small	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	12.32%	21.18%	37.07%	20.78%	31.59%	43.63%	9.39%	25.39%	33.60%	24.50%	46.42%	54.82%
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	13.57%	22.45%	38.57%	22.49%	33.10%	45.93%	9.64%	26.93%	35.48%	26.10%	48.98%	62.80%
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	12.11%	20.36%	36.30%	20.15%	30.98%	39.71%	8.98%	25.54%	34.01%	24.93%	47.38%	57.60%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	5.49%	13.45%	27.05%	12.25%	21.41%	40.05%	7.15%	18.91%	25.61%	15.46%	40.36%	55.74%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	12.34%	21.17%	36.84%	21.22%	31.69%	43.76%	9.46%	25.69%	34.03%	24.80%	47.05%	58.03%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	13.22%	22.15%	37.98%	22.16%	32.62%	44.84%	9.69%	26.52%	35.00%	25.79%	47.79%	59.44%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	11.86%	20.62%	36.33%	20.48%	30.72%	41.38%	9.18%	25.28%	33.63%	24.59%	46.50%	56.97%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	4.56%	12.50%	25.43%	11.37%	20.31%	37.12%	6.93%	18.22%	24.88%	14.78%	39.27%	51.30%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	13.00%	21.91%	37.69%	22.02%	32.40%	44.12%	9.72%	26.35%	34.79%	25.61%	47.64%	58.60%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	12.34%	21.46%	37.58%	21.20%	32.04%	44.01%	9.42%	25.69%	33.95%	24.84%	46.15%	59.71%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	13.04%	21.91%	37.74%	22.03%	32.37%	44.10%	9.71%	26.35%	34.78%	25.63%	47.61%	58.95%
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	17.02%	26.38%	43.90%	26.87%	38.30%	51.16%	11.04%	30.44%	39.64%	30.48%	53.28%	60.47%
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	16.88%	26.26%	43.86%	26.73%	38.24%	51.16%	11.03%	30.28%	39.41%	30.26%	53.15%	60.35%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	4.28%	12.36%	24.95%	11.22%	19.79%	41.45%	7.12%	18.61%	25.37%	15.03%	40.73%	53.34%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	4.28%	12.36%	24.95%	11.22%	19.79%	41.45%	7.12%	18.61%	25.37%	15.03%	40.73%	53.34%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	4.28%	12.36%	24.95%	11.22%	19.79%	41.45%	7.12%	18.61%	25.37%	15.03%	40.73%	53.34%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	10.52%	20.27%	36.67%	19.88%	31.31%	45.61%	9.55%	24.99%	33.02%	23.30%	47.30%	58.91%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	11.31%	21.05%	38.46%	20.32%	32.42%	49.24%	9.66%	25.82%	34.27%	24.60%	48.72%	64.28%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	11.31%	21.05%	38.46%	20.32%	32.42%	49.24%	9.66%	25.82%	34.27%	24.60%	48.72%	64.28%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	10.78%	20.52%	37.80%	20.06%	32.03%	48.10%	9.67%	25.07%	33.36%	23.74%	47.77%	63.01%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	4.41%	12.59%	25.37%	11.24%	20.35%	43.48%	7.36%	18.73%	25.75%	15.30%	41.35%	57.65%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	4.44%	12.61%	25.42%	11.28%	20.37%	43.44%	7.38%	18.72%	25.74%	15.28%	41.36%	57.78%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	10.78%	20.52%	37.80%	20.06%	32.03%	48.10%	9.67%	25.07%	33.36%	23.74%	47.77%	63.01%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	4.41%	12.59%	25.38%	11.24%	20.35%	43.48%	7.36%	18.73%	25.75%	15.30%	41.35%	57.65%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	4.44%	12.61%	25.42%	11.28%	20.37%	43.44%	7.38%	18.72%	25.74%	15.28%	41.36%	57.78%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	10.78%	20.52%	37.80%	20.06%	32.03%	48.10%	9.67%	25.07%	33.36%	23.74%	47.77%	63.01%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	11.31%	21.05%	38.46%	20.32%	32.42%	49.24%	9.66%	25.82%	34.27%	24.60%	48.72%	64.28%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	11.31%	21.05%	38.46%	20.32%	32.42%	49.24%	9.66%	25.82%	34.27%	24.60%	48.72%	64.28%
Medium	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	6.34%	15.57%	30.09%	14.68%	24.96%	47.21%	7.97%	21.31%	29.26%	19.65%	43.77%	59.49%
Medium	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	6.34%	15.57%	30.09%	14.68%	24.96%	47.21%	7.97%	21.31%	29.26%	19.65%	43.77%	59.49%
Medium	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	11.63%	21.33%	38.71%	20.91%	32.62%	47.69%	9.89%	26.13%	34.29%	24.88%	48.17%	59.66%
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	17.33%	26.75%	44.54%	27.12%	38.94%	51.03%	11.20%	30.87%	39.98%	30.87%	53.73%	63.44%
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	10.30%	19.86%	36.79%	19.25%	31.23%	47.98%	9.38%	25.46%	33.71%	23.88%	49.05%	60.96%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	5.86%	14.49%	28.81%	13.16%	23.24%	46.44%	7.90%	20.67%	27.99%	17.41%	44.15%	57.24%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	12.85%	22.72%	40.45%	22.40%	34.55%	49.59%	10.07%	27.21%	35.74%	26.19%	49.83%	60.55%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	17.04%	26.38%	43.91%	26.88%	38.40%	50.59%	11.09%	30.54%	39.63%	30.48%	53.42%	59.62%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	11.18%	20.79%	38.10%	20.33%	31.86%	49.39%	9.49%	25.48%	33.79%	24.44%	47.93%	60.70%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	4.92%	12.68%	25.85%	11.16%	20.29%	43.76%	7.21%	19.10%	26.13%	15.84%	41.83%	56.14%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	16.64%	26.06%	43.57%	26.39%	38.14%	50.10%	10.99%	30.27%	39.24%	29.94%	53.15%	59.50%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	16.16%	25.68%	43.51%	25.78%	37.79%	50.34%	10.80%	29.70%	38.73%	29.71%	51.88%	58.99%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	16.67%	26.07%	43.68%	26.38%	38.12%	50.02%	10.99%	30.27%	39.26%	29.96%	53.13%	59.46%

1280 — 1280×1280, full-frame weights

Size	Platform	Runtime	Decode	APs	Δ APs vs FP32 (pp)	AP	AP50	Time per image (ms)	Realized FPS	Source
Nano	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	19.75%	ref	28.44%	46.70%	22.59	136.8	v-1974
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	19.62%	-0.13	28.25%	46.56%	16.48	177.1	v-19ce
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	4.96%	-14.79	13.17%	25.76%	1486.28	1.1	v-1c38
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	4.96%	-14.79	13.17%	25.76%	604.04	1.7	v-1b3f
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	4.96%	-14.79	13.17%	25.76%	604.09	1.7	v-1b45
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	16.29%	-3.46	24.50%	41.69%	615.95	1.8	v-1b4d
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	17.73%	-2.02	26.08%	44.01%	231.48	6.2	v-1765
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	17.73%	-2.02	26.08%	44.01%	340.59	16.4	v-1a10
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	17.46%	-2.29	25.87%	44.00%	211.45	6.1	v-1af0
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	3.66%	-16.09	12.38%	24.92%	279.95	3.8	v-1aed
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	3.66%	-16.09	12.38%	24.92%	274.50	3.9	v-1ae9
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	17.45%	-2.30	25.87%	44.00%	282.15	17.3	v-1af2
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	3.66%	-16.09	12.38%	24.92%	307.60	14.0	v-1aee
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	3.66%	-16.09	12.38%	24.92%	306.35	14.1	v-1aec
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	17.45%	-2.30	25.87%	44.00%	280.11	17.4	v-188e
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	17.73%	-2.02	26.08%	44.01%	240.10	6.1	v-1785
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	17.72%	-2.03	26.08%	44.01%	360.68	15.4	v-1a19
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	18.61%	-1.14	27.19%	45.52%	81.77	15.0	v-162f
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	19.70%	-0.05	28.31%	46.68%	78.04	70.9	v-1914
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	16.37%	-3.38	25.21%	42.89%	69.45	91.6	v-1915
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	6.01%	-13.74	14.95%	28.88%	24.98	147.1	v-166e
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	18.55%	-1.20	26.99%	45.48%	29.11	124.6	v-166f
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	19.59%	-0.16	28.21%	46.44%	61.37	56.7	v-1670
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	17.75%	-2.00	26.19%	44.54%	29.89	126.3	v-1d56
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	6.26%	-13.49	14.99%	28.99%	19.16	149.5	v-1d57
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	19.55%	-0.20	28.18%	46.39%	62.14	56.0	v-1d58
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	19.19%	-0.56	28.00%	46.20%	47.05	51.7	v-1d5a
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	19.17%	-0.58	27.98%	46.14%	48.23	50.9	v-1d59
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	23.91%	ref	32.86%	52.89%	42.34	101.6	v-1979
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	23.66%	-0.25	32.65%	52.80%	22.29	146.0	v-19d2
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	5.81%	-18.10	14.77%	30.08%	2073.00	0.9	v-1c59
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	5.81%	-18.10	14.77%	30.08%	1144.38	0.9	v-1b85
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	5.81%	-18.10	14.77%	30.08%	1144.09	0.9	v-1b8f
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	20.96%	-2.95	28.86%	48.34%	1154.48	0.9	v-1b9d
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	21.86%	-2.05	30.29%	50.34%	420.96	2.9	v-177e
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	21.86%	-2.05	30.29%	50.34%	617.03	7.6	v-1a1f
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	21.90%	-2.01	30.16%	50.34%	449.13	2.5	v-1b31
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	6.37%	-17.54	15.59%	31.32%	492.25	2.1	v-1b27
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	6.37%	-17.54	15.59%	31.32%	497.38	2.1	v-1b20
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	21.90%	-2.01	30.16%	50.34%	570.93	7.7	v-1b32
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	6.37%	-17.54	15.59%	31.32%	589.70	7.1	v-1b2e
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	6.37%	-17.54	15.59%	31.32%	582.04	7.2	v-1b25
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	21.90%	-2.01	30.16%	50.34%	573.82	7.7	v-18a4
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	21.86%	-2.05	30.29%	50.34%	431.04	2.9	v-179a
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	21.86%	-2.05	30.29%	50.34%	627.33	7.6	v-1a28
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	23.70%	-0.21	32.69%	52.86%	116.20	41.6	v-15cd
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	20.23%	-3.68	26.42%	46.69%	80.72	64.5	v-15ce
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	7.59%	-16.32	17.15%	33.15%	28.93	90.7	v-16a3
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	23.11%	-0.80	31.88%	52.08%	39.42	97.9	v-16a4
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	24.11%	+0.20	32.97%	53.05%	81.65	39.1	v-16a7
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	22.35%	-1.56	30.71%	51.14%	39.42	99.3	v-1d7e
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	7.17%	-16.74	16.65%	32.95%	28.67	91.4	v-1d7f
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	24.03%	+0.12	32.95%	52.97%	82.26	38.9	v-1d80
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	23.44%	-0.47	32.55%	52.66%	61.95	37.0	v-1d82
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	24.02%	+0.11	32.91%	52.99%	62.71	36.9	v-1d81
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	27.73%	ref	36.42%	57.89%	83.97	54.0	v-1981
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	27.44%	-0.29	36.18%	57.83%	56.38	86.7	v-19d6
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	7.25%	-20.48	17.02%	33.26%	2814.51	0.6	v-1c77
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	7.25%	-20.48	17.02%	33.26%	1791.17	0.6	v-1c04
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	7.25%	-20.48	17.02%	33.26%	1788.25	0.6	v-1c1a
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	23.77%	-3.96	32.40%	53.64%	1800.30	0.6	v-1c34
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	24.93%	-2.80	34.03%	55.80%	693.42	1.7	v-17c3
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	24.93%	-2.80	34.03%	55.80%	1560.19	2.7	v-1a30
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	24.80%	-2.93	33.74%	55.68%	670.86	1.6	v-1b79
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	7.65%	-20.08	17.62%	34.53%	746.50	1.4	v-1b6b
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	7.65%	-20.08	17.62%	34.53%	745.40	1.4	v-1b61
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	24.80%	-2.93	33.74%	55.68%	1521.10	2.7	v-1b80
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	7.64%	-20.09	17.62%	34.53%	1501.13	2.7	v-1b73
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	7.65%	-20.08	17.62%	34.53%	1500.19	2.7	v-1b69
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	24.80%	-2.93	33.74%	55.68%	1521.82	2.7	v-18ae
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	24.93%	-2.80	34.03%	55.80%	695.36	1.7	v-17e4
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	24.93%	-2.80	34.03%	55.80%	1565.67	2.7	v-1a33
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	27.42%	-0.31	36.24%	57.86%	209.00	21.3	v-15de
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	23.29%	-4.44	31.76%	53.33%	127.36	36.3	v-15e2
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	7.86%	-19.87	18.30%	34.99%	45.37	51.2	v-16e5
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	26.74%	-0.99	35.31%	56.80%	60.63	53.6	v-16e6
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	27.78%	+0.05	36.50%	58.06%	129.62	20.7	v-16e7
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	25.17%	-2.56	34.10%	56.03%	57.92	56.3	v-1db7
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	7.77%	-19.96	17.79%	34.72%	43.75	54.1	v-1db8
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	27.77%	+0.04	36.48%	58.03%	132.06	20.2	v-1db9
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	27.20%	-0.53	36.13%	57.79%	107.51	20.0	v-1dbb
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	27.20%	-0.53	36.14%	57.76%	111.28	19.6	v-1dba

Sizes	Platform	Runtime	Decode	Status
Nano, Small, Medium	NXP Ara240 (FRDM)	Ara DVM	Smart	Not supported: ara2-converter 2.7.3: ARA-240 area attention spans at most 1024 positions; 1280x1280 spans 1600 (DE-2996)
Small, Medium	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	Not supported: Hailo-8L: insufficient on-chip memory for this model and input size

Full metrics — 1280

Size	Platform	Runtime	Decode	Precision	APs	AP	AP50	AP75	APm	APl	AR@1	AR@10	AR@100	ARs	ARm	ARl
Nano	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	19.75%	28.44%	46.70%	28.73%	39.68%	49.07%	11.49%	32.88%	43.78%	34.94%	56.67%	66.04%
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	19.62%	28.25%	46.56%	28.57%	39.49%	48.93%	11.45%	32.65%	43.53%	34.72%	56.38%	66.04%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	4.96%	13.17%	25.76%	12.47%	22.33%	38.47%	7.67%	20.23%	27.83%	17.50%	42.30%	54.42%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	4.96%	13.17%	25.76%	12.47%	22.33%	38.47%	7.67%	20.23%	27.83%	17.50%	42.30%	54.42%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	4.96%	13.17%	25.76%	12.47%	22.33%	38.47%	7.67%	20.23%	27.83%	17.50%	42.30%	54.42%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	16.29%	24.50%	41.69%	24.73%	34.81%	42.43%	10.44%	29.90%	39.87%	30.94%	51.49%	59.78%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	17.73%	26.08%	44.01%	26.57%	36.96%	49.85%	10.82%	31.12%	41.83%	33.05%	54.57%	65.81%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	17.73%	26.08%	44.01%	26.57%	36.96%	49.52%	10.82%	31.12%	41.83%	33.05%	54.57%	65.81%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	17.46%	25.87%	44.00%	26.16%	36.70%	46.52%	10.92%	31.14%	41.61%	32.85%	54.17%	64.83%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	3.66%	12.38%	24.92%	11.16%	23.56%	42.95%	7.29%	18.96%	26.28%	15.59%	41.76%	57.29%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	3.66%	12.38%	24.92%	11.16%	23.56%	42.95%	7.29%	18.96%	26.28%	15.59%	41.76%	57.29%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	17.45%	25.87%	44.00%	26.16%	36.70%	46.52%	10.92%	31.14%	41.61%	32.85%	54.17%	64.83%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	3.66%	12.38%	24.92%	11.16%	23.56%	42.95%	7.29%	18.96%	26.28%	15.59%	41.76%	57.29%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	3.66%	12.38%	24.92%	11.16%	23.56%	42.95%	7.29%	18.96%	26.28%	15.59%	41.76%	57.29%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	17.45%	25.87%	44.00%	26.16%	36.70%	46.11%	10.92%	31.14%	41.61%	32.85%	54.17%	64.83%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	17.73%	26.08%	44.01%	26.57%	36.96%	49.85%	10.82%	31.12%	41.83%	33.05%	54.57%	65.81%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	17.72%	26.08%	44.01%	26.57%	36.97%	49.85%	10.82%	31.12%	41.83%	33.05%	54.57%	65.81%
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	18.61%	27.19%	45.52%	27.53%	38.08%	48.80%	11.12%	31.58%	42.20%	33.53%	54.52%	64.60%
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	19.70%	28.31%	46.68%	28.76%	39.59%	49.02%	11.46%	32.72%	43.61%	34.76%	56.52%	66.91%
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	16.37%	25.21%	42.89%	25.64%	36.57%	50.66%	10.76%	29.80%	39.81%	31.10%	52.62%	63.62%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	6.01%	14.95%	28.88%	14.12%	25.13%	46.69%	8.25%	21.42%	30.16%	20.46%	44.88%	60.76%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	18.55%	26.99%	45.48%	27.32%	37.87%	48.95%	10.99%	31.60%	42.38%	33.65%	55.01%	65.49%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	19.59%	28.21%	46.44%	28.72%	39.37%	48.99%	11.49%	32.92%	43.77%	34.97%	56.40%	66.37%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	17.75%	26.19%	44.54%	26.42%	36.91%	47.51%	11.00%	31.14%	41.62%	32.70%	54.57%	65.52%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	6.26%	14.99%	28.99%	14.01%	24.95%	43.62%	8.18%	21.54%	30.12%	20.30%	45.20%	59.71%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	19.55%	28.18%	46.39%	28.63%	39.38%	49.18%	11.44%	32.84%	43.76%	34.92%	56.52%	65.92%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	19.19%	28.00%	46.20%	28.43%	39.29%	48.81%	11.39%	32.63%	43.31%	34.47%	55.65%	65.87%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	19.17%	27.98%	46.14%	28.38%	39.35%	48.90%	11.40%	32.55%	43.37%	34.51%	55.76%	65.11%
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	23.91%	32.86%	52.89%	34.20%	44.45%	54.52%	12.62%	36.46%	48.05%	39.81%	60.32%	74.24%
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	23.66%	32.65%	52.80%	33.78%	44.25%	54.17%	12.55%	36.22%	47.80%	39.58%	59.98%	74.46%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	5.81%	14.77%	30.08%	13.08%	26.11%	43.53%	7.99%	21.52%	29.88%	19.46%	45.33%	60.88%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	5.81%	14.77%	30.08%	13.08%	26.11%	43.53%	7.99%	21.52%	29.88%	19.46%	45.33%	60.88%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	5.81%	14.77%	30.08%	13.08%	26.11%	43.53%	7.99%	21.52%	29.88%	19.46%	45.33%	60.88%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	20.96%	28.86%	48.34%	29.71%	39.42%	48.42%	11.54%	33.77%	45.02%	37.14%	56.77%	68.31%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	21.86%	30.29%	50.34%	31.12%	41.29%	50.37%	12.08%	35.01%	46.52%	38.40%	58.60%	68.52%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	21.86%	30.29%	50.34%	31.12%	41.29%	50.37%	12.08%	35.01%	46.52%	38.40%	58.60%	68.52%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	21.90%	30.16%	50.34%	30.94%	40.99%	47.78%	11.95%	34.94%	46.31%	38.36%	58.09%	67.96%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	6.37%	15.59%	31.32%	14.13%	26.59%	45.86%	8.39%	22.70%	31.31%	20.42%	47.75%	63.30%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	6.37%	15.59%	31.32%	14.13%	26.59%	45.86%	8.39%	22.70%	31.31%	20.42%	47.75%	63.30%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	21.90%	30.16%	50.34%	30.94%	40.99%	47.78%	11.95%	34.94%	46.31%	38.36%	58.09%	67.96%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	6.37%	15.59%	31.32%	14.13%	26.59%	45.86%	8.39%	22.70%	31.31%	20.42%	47.75%	63.30%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	6.37%	15.59%	31.32%	14.13%	26.59%	45.86%	8.39%	22.70%	31.31%	20.42%	47.75%	63.30%
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	21.90%	30.16%	50.34%	30.94%	40.99%	47.78%	11.95%	34.94%	46.31%	38.36%	58.09%	67.96%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	21.86%	30.29%	50.34%	31.12%	41.29%	50.37%	12.08%	35.01%	46.52%	38.40%	58.60%	68.52%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	21.86%	30.29%	50.34%	31.12%	41.29%	50.37%	12.08%	35.01%	46.52%	38.40%	58.60%	68.52%
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	23.70%	32.69%	52.86%	33.81%	44.34%	54.06%	12.64%	36.25%	47.75%	39.42%	60.09%	73.99%
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	20.23%	26.42%	46.69%	26.12%	36.86%	50.29%	10.77%	32.20%	43.74%	36.66%	54.71%	64.74%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	7.59%	17.15%	33.15%	15.72%	28.06%	47.43%	8.82%	23.87%	33.22%	23.34%	48.18%	66.15%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	23.11%	31.88%	52.08%	32.83%	43.06%	52.23%	12.34%	35.87%	47.28%	39.12%	59.28%	71.79%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	24.11%	32.97%	53.05%	34.09%	44.48%	53.51%	12.78%	36.55%	48.17%	40.09%	60.11%	75.15%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	22.35%	30.71%	51.14%	31.72%	41.67%	53.92%	11.89%	35.09%	46.48%	38.41%	58.46%	71.40%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	7.17%	16.65%	32.95%	15.00%	28.08%	46.84%	8.48%	23.11%	32.19%	22.24%	47.02%	62.67%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	24.03%	32.95%	52.97%	34.05%	44.52%	52.69%	12.81%	36.52%	48.12%	39.96%	60.16%	76.10%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	23.44%	32.55%	52.66%	33.39%	44.38%	52.83%	12.65%	36.05%	47.59%	39.40%	59.46%	73.88%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	24.02%	32.91%	52.99%	34.06%	44.45%	52.56%	12.80%	36.49%	48.08%	39.93%	60.15%	75.99%
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	27.73%	36.42%	57.89%	38.04%	48.25%	55.54%	13.68%	39.69%	51.44%	43.76%	62.77%	73.20%
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	27.44%	36.18%	57.83%	37.76%	48.03%	55.45%	13.60%	39.46%	51.18%	43.46%	62.54%	73.19%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	7.25%	17.02%	33.26%	15.86%	27.80%	45.55%	8.95%	23.19%	31.86%	21.85%	47.04%	59.53%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	7.25%	17.02%	33.26%	15.86%	27.81%	45.55%	8.95%	23.19%	31.86%	21.85%	47.04%	59.53%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	7.25%	17.02%	33.26%	15.86%	27.81%	45.55%	8.95%	23.19%	31.86%	21.85%	47.04%	59.53%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	23.77%	32.40%	53.64%	33.19%	43.90%	50.99%	12.69%	35.98%	47.13%	39.40%	58.40%	68.09%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	24.93%	34.03%	55.80%	35.18%	46.17%	53.58%	13.21%	37.40%	48.86%	41.13%	60.36%	70.16%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	24.93%	34.03%	55.80%	35.18%	46.17%	53.58%	13.21%	37.40%	48.86%	41.13%	60.36%	70.16%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	24.80%	33.74%	55.68%	34.61%	45.53%	51.95%	13.06%	37.12%	48.47%	40.93%	59.76%	68.81%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	7.65%	17.62%	34.53%	16.06%	28.70%	46.78%	9.24%	24.12%	32.97%	22.96%	48.25%	63.84%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	7.65%	17.62%	34.53%	16.06%	28.70%	46.78%	9.24%	24.12%	32.97%	22.96%	48.25%	63.84%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	24.80%	33.74%	55.68%	34.61%	45.53%	51.95%	13.06%	37.12%	48.47%	40.93%	59.76%	68.81%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	7.64%	17.62%	34.53%	16.06%	28.70%	46.78%	9.24%	24.12%	32.97%	22.96%	48.25%	63.84%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	7.65%	17.62%	34.53%	16.06%	28.70%	46.77%	9.24%	24.12%	32.97%	22.96%	48.25%	63.84%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	24.80%	33.74%	55.68%	34.61%	45.53%	51.95%	13.06%	37.12%	48.47%	40.93%	59.76%	68.81%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	24.93%	34.03%	55.80%	35.18%	46.17%	53.58%	13.21%	37.40%	48.86%	41.13%	60.36%	70.16%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	24.93%	34.03%	55.80%	35.18%	46.17%	53.58%	13.21%	37.40%	48.86%	41.13%	60.36%	70.16%
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	27.42%	36.24%	57.86%	37.87%	48.20%	55.60%	13.64%	39.54%	51.26%	43.45%	62.74%	73.69%
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	23.29%	31.76%	53.33%	33.06%	43.69%	53.26%	12.56%	36.37%	48.02%	40.25%	59.32%	70.72%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	7.86%	18.30%	34.99%	17.15%	30.61%	50.10%	9.50%	24.87%	34.30%	24.18%	49.58%	65.16%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	26.74%	35.31%	56.80%	36.97%	47.02%	53.06%	13.41%	38.40%	50.17%	42.63%	61.45%	72.53%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	27.78%	36.50%	58.06%	38.21%	48.35%	54.02%	13.70%	39.57%	51.49%	43.94%	62.64%	73.65%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	25.17%	34.10%	56.03%	35.13%	46.13%	51.88%	13.14%	37.44%	48.97%	41.24%	60.67%	70.34%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	7.77%	17.79%	34.72%	16.47%	29.58%	48.56%	9.19%	24.45%	33.35%	23.38%	48.74%	65.13%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	27.77%	36.48%	58.03%	38.13%	48.22%	53.97%	13.75%	39.51%	51.34%	43.79%	62.47%	73.37%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	27.20%	36.13%	57.79%	37.62%	48.17%	54.58%	13.62%	39.08%	50.86%	43.21%	61.95%	73.59%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	27.20%	36.14%	57.76%	37.77%	48.16%	54.42%	13.64%	39.08%	50.80%	43.12%	61.95%	73.25%

1344x768 — 1344×768, tiled weights

Size	Platform	Runtime	Decode	APs	Δ APs vs FP32 (pp)	AP	AP50	Time per image (ms)	Realized FPS	Source
Nano	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	20.70%	ref	29.95%	48.59%	17.68	179.0	v-1976
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	20.53%	-0.17	29.79%	48.46%	13.46	243.1	v-19d0
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	6.44%	-14.26	15.17%	28.75%	891.37	1.8	v-1c49
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	6.44%	-14.26	15.17%	28.75%	328.97	3.2	v-1b6a
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	6.44%	-14.26	15.17%	28.75%	328.86	3.2	v-1b6d
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	16.78%	-3.92	25.48%	42.94%	339.16	3.4	v-1b72
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	18.57%	-2.13	27.40%	46.04%	152.51	12.2	v-1778
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	18.57%	-2.13	27.40%	46.04%	213.35	23.5	v-1a1c
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	18.26%	-2.44	27.23%	45.86%	121.18	12.0	v-1b0e
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	7.16%	-13.54	16.03%	30.70%	153.21	7.2	v-1b09
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	7.16%	-13.54	16.04%	30.72%	158.15	7.0	v-1b06
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	18.26%	-2.44	27.23%	45.86%	175.64	28.0	v-1b10
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	7.16%	-13.54	16.03%	30.70%	187.44	24.1	v-1b0b
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	7.16%	-13.54	16.04%	30.72%	184.70	24.2	v-1b08
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	18.26%	-2.44	27.23%	45.85%	175.80	27.8	v-18a2
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	18.57%	-2.13	27.40%	46.04%	159.67	12.1	v-1790
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	18.57%	-2.13	27.40%	46.04%	220.79	21.4	v-1a26
Nano	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	11.58%	-9.12	21.45%	37.28%	38.10	48.5	v-16db
Nano	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	11.58%	-9.12	21.44%	37.28%	32.66	100.0	v-192e
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	19.10%	-1.60	28.26%	46.99%	50.34	25.5	v-1636
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	20.48%	-0.22	29.80%	48.47%	69.33	99.2	v-1919
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	17.86%	-2.84	27.02%	45.23%	65.79	109.7	v-191a
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	7.75%	-12.95	17.80%	32.87%	14.83	235.5	v-1683
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	19.75%	-0.95	28.82%	47.58%	20.89	155.5	v-1684
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	20.71%	+0.01	29.93%	48.54%	39.20	91.2	v-1686
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	18.77%	-1.93	27.80%	46.59%	20.92	150.4	v-1d60
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	7.35%	-13.35	16.77%	31.71%	16.16	237.7	v-1d61
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	20.67%	-0.03	29.89%	48.57%	40.04	89.5	v-1d62
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	20.35%	-0.35	29.62%	48.42%	33.12	83.7	v-1d64
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	20.35%	-0.35	29.63%	48.44%	32.72	81.0	v-1d63
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	25.45%	ref	34.62%	55.38%	23.65	137.8	v-197e
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	25.27%	-0.18	34.46%	55.28%	15.57	195.0	v-19d4
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	7.60%	-17.85	16.74%	32.49%	1225.10	1.6	v-1c6d
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	7.60%	-17.85	16.75%	32.49%	604.67	1.7	v-1bce
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	7.60%	-17.85	16.75%	32.48%	604.78	1.7	v-1bd9
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	19.62%	-5.83	28.32%	49.03%	613.03	1.7	v-1be8
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	20.00%	-5.45	29.02%	50.37%	229.15	5.8	v-1791
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	20.00%	-5.45	29.01%	50.37%	396.68	12.0	v-1a24
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	19.68%	-5.77	28.73%	50.02%	226.97	5.1	v-1b50
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	7.91%	-17.54	17.45%	33.84%	260.29	4.0	v-1b4c
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	7.98%	-17.47	17.54%	33.96%	267.02	3.9	v-1b48
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	19.68%	-5.77	28.73%	50.03%	356.89	12.2	v-1b51
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	7.91%	-17.54	17.44%	33.83%	351.08	11.8	v-1b4e
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	7.98%	-17.47	17.52%	33.96%	339.45	12.2	v-1b4a
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	19.68%	-5.77	28.73%	50.02%	358.44	12.2	v-18a7
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	20.00%	-5.45	29.02%	50.37%	239.23	5.8	v-17c2
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	20.00%	-5.45	29.02%	50.37%	405.52	12.0	v-1a2e
Small	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	13.54%	-11.91 ⚠	23.32%	40.98%	54.14	26.4	v-1624
Small	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	13.54%	-11.91 ⚠	23.32%	40.98%	45.36	33.4	v-1950
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	25.19%	-0.26	34.45%	55.21%	84.26	62.1	v-15d7
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	21.51%	-3.94	27.55%	48.53%	67.15	93.2	v-15da
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	9.88%	-15.57	20.12%	37.56%	21.68	139.9	v-16a9
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	23.10%	-2.35	32.53%	53.36%	31.71	145.5	v-16aa
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	25.46%	+0.01	34.56%	55.25%	52.61	60.8	v-16ab
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	20.18%	-5.27	29.31%	50.77%	27.20	147.6	v-1d9e
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	8.33%	-17.12	18.11%	35.07%	20.19	138.5	v-1d9f
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	25.34%	-0.11	34.53%	55.28%	52.53	60.6	v-1da0
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	25.11%	-0.34	34.24%	55.10%	43.16	55.7	v-1da2
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	25.15%	-0.30	34.33%	55.17%	43.40	55.9	v-1da1
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	29.36%	ref	37.81%	59.76%	57.33	79.8	v-1988
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	29.09%	-0.27	37.58%	59.57%	34.77	126.5	v-19da
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	3.01%	-26.35	8.79%	16.59%	1717.95	1.0	v-1c86
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	3.01%	-26.35	8.79%	16.59%	1037.91	1.0	v-1c5d
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	3.00%	-26.36	8.78%	16.56%	1037.96	1.0	v-1c65
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	7.78%	-21.58 ⚠	14.20%	23.90%	1045.41	1.0	v-1c6b
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	20.85%	-8.51	30.63%	51.99%	409.90	3.1	v-17b5
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	20.85%	-8.51	30.63%	51.99%	990.13	4.3	v-1a35
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	20.75%	-8.61	30.32%	51.88%	376.57	2.9	v-1bb5
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	7.92%	-21.44	17.52%	34.36%	419.08	2.5	v-1baf
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	7.92%	-21.44	17.52%	34.36%	425.29	2.4	v-1ba8
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	20.75%	-8.61	30.32%	51.88%	948.51	4.4	v-1bb8
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	7.92%	-21.44	17.52%	34.36%	926.41	4.4	v-1bb0
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	7.92%	-21.44	17.51%	34.36%	926.49	4.4	v-1bab
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	20.75%	-8.61	30.32%	51.88%	948.75	4.4	v-18c2
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	20.85%	-8.51	30.63%	51.99%	411.06	3.1	v-17e9
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	20.85%	-8.51	30.63%	51.99%	983.94	4.4	v-1a3b
Medium	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	12.55%	-16.81 ⚠	22.58%	42.67%	127.82	9.1	v-163d
Medium	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	12.55%	-16.81 ⚠	22.58%	42.67%	119.78	9.8	v-1957
Medium	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	18.64%	-10.72 ⚠	28.79%	50.43%	153.55	7.2	v-1653
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	29.05%	-0.31	37.57%	59.59%	141.68	32.5	v-195e
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	18.05%	-11.31 ⚠	24.92%	46.19%	96.87	51.7	v-1961
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	9.18%	-20.18	19.68%	38.28%	30.40	82.6	v-16fa
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	27.52%	-1.84	36.20%	58.27%	37.83	85.4	v-16fc
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	29.57%	+0.21	37.97%	59.92%	85.27	31.5	v-16fd
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	21.96%	-7.40	31.85%	52.65%	37.66	87.5	v-1ddf
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	7.38%	-21.98	16.65%	32.49%	29.99	84.3	v-1de0
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	29.47%	+0.11	37.91%	59.85%	88.53	30.3	v-1de1
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	29.17%	-0.19	37.66%	59.73%	72.69	30.4	v-1de3
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	28.77%	-0.59	37.38%	59.44%	76.00	29.4	v-1de2

Sizes	Platform	Runtime	Decode	Status
Small	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	Not supported: Hailo-8L: insufficient on-chip memory for this model and input size

Full metrics — 1344x768

Size	Platform	Runtime	Decode	Precision	APs	AP	AP50	AP75	APm	APl	AR@1	AR@10	AR@100	ARs	ARm	ARl
Nano	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	20.70%	29.95%	48.59%	30.68%	41.50%	50.94%	11.74%	33.98%	45.03%	35.84%	58.08%	66.57%
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	20.53%	29.79%	48.46%	30.53%	41.39%	50.91%	11.70%	33.81%	44.82%	35.56%	57.97%	66.51%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	6.44%	15.17%	28.75%	14.54%	24.59%	39.36%	8.19%	21.69%	30.21%	20.51%	43.49%	53.12%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	6.44%	15.17%	28.75%	14.54%	24.59%	39.36%	8.19%	21.69%	30.21%	20.51%	43.49%	53.12%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	6.44%	15.17%	28.75%	14.54%	24.59%	39.36%	8.19%	21.69%	30.21%	20.51%	43.49%	53.12%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	16.78%	25.48%	42.94%	26.04%	35.76%	42.76%	10.67%	30.16%	40.28%	31.58%	51.46%	57.00%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	18.57%	27.40%	46.04%	28.17%	38.38%	46.43%	11.17%	31.61%	42.15%	33.60%	54.20%	60.88%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	18.57%	27.40%	46.04%	28.17%	38.38%	46.43%	11.17%	31.61%	42.15%	33.60%	54.20%	60.88%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	18.26%	27.23%	45.86%	27.68%	38.32%	47.41%	11.17%	31.25%	41.93%	33.12%	54.35%	64.15%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	7.16%	16.03%	30.70%	14.92%	26.11%	43.02%	8.52%	22.51%	31.53%	21.78%	45.84%	56.28%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	7.16%	16.04%	30.72%	14.93%	26.12%	43.04%	8.53%	22.53%	31.52%	21.77%	45.85%	56.28%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	18.26%	27.23%	45.86%	27.68%	38.32%	47.41%	11.17%	31.25%	41.93%	33.12%	54.35%	64.15%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	7.16%	16.03%	30.70%	14.92%	26.12%	43.02%	8.52%	22.51%	31.53%	21.78%	45.84%	56.28%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	7.16%	16.04%	30.72%	14.93%	26.12%	43.04%	8.53%	22.53%	31.52%	21.77%	45.85%	56.28%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	18.26%	27.23%	45.85%	27.68%	38.31%	47.41%	11.17%	31.25%	41.93%	33.12%	54.35%	64.15%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	18.57%	27.40%	46.04%	28.17%	38.38%	46.43%	11.17%	31.61%	42.15%	33.60%	54.20%	60.88%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	18.57%	27.40%	46.04%	28.17%	38.38%	46.43%	11.17%	31.61%	42.15%	33.60%	54.20%	60.88%
Nano	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	11.58%	21.45%	37.28%	21.87%	33.72%	44.25%	9.83%	26.60%	36.99%	28.34%	48.86%	60.21%
Nano	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	11.58%	21.44%	37.28%	21.87%	33.72%	44.25%	9.83%	26.60%	36.99%	28.34%	48.86%	60.21%
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	19.10%	28.26%	46.99%	28.85%	39.76%	46.23%	11.42%	32.26%	43.13%	34.13%	55.91%	64.32%
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	20.48%	29.80%	48.47%	30.60%	41.47%	50.92%	11.70%	33.87%	44.84%	35.58%	58.00%	66.37%
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	17.86%	27.02%	45.23%	27.78%	38.24%	49.25%	10.97%	31.45%	41.95%	32.90%	54.39%	63.61%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	7.75%	17.80%	32.87%	17.41%	29.29%	46.78%	8.96%	23.96%	33.27%	23.12%	48.09%	61.28%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	19.75%	28.82%	47.58%	29.33%	40.06%	50.94%	11.50%	33.01%	43.92%	34.78%	56.79%	66.65%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	20.71%	29.93%	48.54%	30.67%	41.62%	48.11%	11.86%	34.09%	45.03%	35.97%	57.89%	65.21%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	18.77%	27.80%	46.59%	28.38%	39.10%	47.52%	11.22%	31.76%	42.33%	33.46%	54.89%	63.00%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	7.35%	16.77%	31.71%	16.07%	27.41%	45.77%	8.60%	23.12%	32.02%	22.07%	46.87%	60.35%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	20.67%	29.89%	48.57%	30.67%	41.61%	48.37%	11.91%	33.98%	44.94%	35.87%	57.86%	65.48%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	20.35%	29.62%	48.42%	30.33%	41.34%	47.56%	11.74%	33.67%	44.50%	35.50%	56.88%	64.10%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	20.35%	29.63%	48.44%	30.37%	41.42%	47.70%	11.82%	33.60%	44.51%	35.42%	57.07%	64.62%
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	25.45%	34.62%	55.38%	35.62%	46.66%	51.67%	13.44%	38.20%	49.86%	41.17%	62.68%	70.43%
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	25.27%	34.46%	55.28%	35.36%	46.59%	51.65%	13.41%	38.03%	49.64%	40.94%	62.53%	70.76%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	7.60%	16.74%	32.49%	15.70%	27.08%	43.01%	9.08%	24.57%	34.11%	24.21%	48.81%	60.34%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	7.60%	16.75%	32.49%	15.71%	27.10%	43.01%	9.08%	24.57%	34.11%	24.21%	48.81%	60.34%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	7.60%	16.75%	32.48%	15.71%	27.10%	43.01%	9.08%	24.57%	34.11%	24.21%	48.81%	60.34%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	19.62%	28.32%	49.03%	28.45%	38.91%	47.24%	11.59%	33.64%	45.05%	36.63%	56.64%	62.88%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	20.00%	29.02%	50.37%	29.58%	40.44%	47.12%	11.77%	34.12%	45.69%	37.35%	57.56%	65.68%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	20.00%	29.01%	50.37%	29.58%	40.42%	47.12%	11.77%	34.12%	45.69%	37.35%	57.56%	65.68%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	19.68%	28.73%	50.02%	29.22%	40.03%	46.14%	11.90%	33.96%	45.57%	37.44%	57.20%	67.62%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	7.91%	17.45%	33.84%	16.21%	28.74%	44.47%	9.24%	25.17%	34.94%	24.86%	50.20%	61.92%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	7.98%	17.54%	33.96%	16.31%	28.77%	44.57%	9.27%	25.22%	35.05%	25.02%	50.26%	62.01%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	19.68%	28.73%	50.03%	29.22%	40.03%	46.14%	11.90%	33.96%	45.57%	37.44%	57.20%	67.62%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	7.91%	17.44%	33.83%	16.20%	28.72%	44.48%	9.24%	25.17%	34.94%	24.86%	50.20%	61.92%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	7.98%	17.52%	33.96%	16.31%	28.76%	44.57%	9.27%	25.22%	35.05%	25.02%	50.26%	62.01%
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	19.68%	28.73%	50.02%	29.22%	40.03%	46.14%	11.90%	33.96%	45.57%	37.44%	57.20%	67.62%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	20.00%	29.02%	50.37%	29.58%	40.44%	47.12%	11.77%	34.12%	45.69%	37.35%	57.56%	65.68%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	20.00%	29.02%	50.37%	29.58%	40.44%	47.12%	11.77%	34.12%	45.69%	37.35%	57.56%	65.68%
Small	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	13.54%	23.32%	40.98%	24.30%	35.86%	43.92%	10.75%	29.72%	41.07%	32.90%	53.20%	64.31%
Small	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	13.54%	23.32%	40.98%	24.30%	35.86%	43.92%	10.75%	29.72%	41.07%	32.90%	53.20%	64.31%
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	25.19%	34.45%	55.21%	35.47%	46.63%	51.59%	13.37%	37.96%	49.53%	40.86%	62.33%	69.78%
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	21.51%	27.55%	48.53%	27.42%	38.37%	48.16%	11.20%	32.59%	44.13%	37.44%	53.78%	64.29%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	9.88%	20.12%	37.56%	19.42%	32.36%	47.77%	9.78%	26.23%	36.20%	26.46%	51.21%	62.53%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	23.10%	32.53%	53.36%	33.30%	44.73%	50.10%	12.69%	36.36%	47.91%	39.28%	60.51%	67.89%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	25.46%	34.56%	55.25%	35.45%	46.49%	50.58%	13.53%	38.29%	49.94%	41.51%	62.45%	66.84%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	20.18%	29.31%	50.77%	29.71%	40.58%	50.41%	11.99%	34.21%	45.49%	37.28%	57.34%	67.77%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	8.33%	18.11%	35.07%	16.90%	29.16%	46.73%	9.40%	25.36%	34.98%	25.09%	49.87%	61.30%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	25.34%	34.53%	55.28%	35.51%	46.58%	50.43%	13.36%	38.13%	49.85%	41.33%	62.41%	66.82%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	25.11%	34.24%	55.10%	35.28%	46.17%	50.17%	13.41%	37.86%	49.45%	41.14%	61.45%	66.31%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	25.15%	34.33%	55.17%	35.31%	46.34%	50.34%	13.31%	37.85%	49.52%	41.07%	61.76%	66.61%
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	29.36%	37.81%	59.76%	39.80%	49.88%	56.59%	13.98%	40.91%	53.03%	45.35%	65.03%	74.60%
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	29.09%	37.58%	59.57%	39.46%	49.69%	56.39%	13.88%	40.63%	52.71%	45.01%	64.71%	74.32%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	3.01%	8.79%	16.59%	8.58%	14.40%	21.03%	5.82%	15.56%	22.12%	13.06%	32.02%	32.99%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	3.01%	8.79%	16.59%	8.58%	14.40%	21.03%	5.82%	15.56%	22.12%	13.06%	32.02%	32.99%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	3.00%	8.78%	16.56%	8.56%	14.40%	21.02%	5.82%	15.56%	22.12%	13.06%	32.02%	32.99%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	7.78%	14.20%	23.90%	14.58%	22.08%	22.78%	7.51%	21.49%	29.52%	19.82%	39.95%	35.30%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	20.85%	30.63%	51.99%	31.52%	43.63%	51.83%	12.44%	35.32%	46.97%	38.90%	59.41%	66.60%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	20.85%	30.63%	51.99%	31.52%	43.63%	51.83%	12.44%	35.32%	46.97%	38.90%	59.41%	66.60%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	20.75%	30.32%	51.88%	31.35%	43.04%	51.50%	12.20%	34.96%	46.63%	38.77%	58.53%	67.66%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	7.92%	17.52%	34.36%	15.98%	28.65%	45.80%	9.16%	25.02%	34.84%	25.41%	49.95%	59.19%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	7.92%	17.52%	34.36%	15.95%	28.65%	45.81%	9.16%	25.03%	34.84%	25.42%	49.94%	59.19%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	20.75%	30.32%	51.88%	31.35%	43.04%	51.50%	12.20%	34.96%	46.63%	38.77%	58.53%	67.66%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	7.92%	17.52%	34.36%	15.98%	28.65%	45.80%	9.16%	25.02%	34.84%	25.41%	49.95%	59.19%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	7.92%	17.51%	34.36%	15.95%	28.64%	45.81%	9.16%	25.03%	34.84%	25.42%	49.94%	59.19%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	20.75%	30.32%	51.88%	31.35%	43.04%	51.50%	12.20%	34.96%	46.63%	38.77%	58.53%	67.66%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	20.85%	30.63%	51.99%	31.52%	43.63%	51.83%	12.44%	35.32%	46.97%	38.90%	59.41%	66.60%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	20.85%	30.63%	51.99%	31.52%	43.63%	51.83%	12.44%	35.32%	46.97%	38.90%	59.41%	66.60%
Medium	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	12.55%	22.58%	42.67%	21.49%	34.82%	48.30%	10.04%	28.64%	40.24%	31.84%	53.00%	63.27%
Medium	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	12.55%	22.58%	42.67%	21.49%	34.82%	48.30%	10.04%	28.64%	40.24%	31.84%	53.00%	63.27%
Medium	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	18.64%	28.79%	50.43%	28.82%	41.45%	51.50%	11.79%	33.75%	45.60%	37.80%	57.09%	66.92%
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	29.05%	37.57%	59.59%	39.55%	49.60%	56.46%	13.88%	40.56%	52.63%	44.94%	64.51%	74.71%
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	18.05%	24.92%	46.19%	23.69%	38.24%	47.20%	10.73%	31.47%	43.13%	35.93%	54.24%	61.34%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	9.18%	19.68%	38.28%	18.10%	32.94%	50.01%	9.80%	26.59%	36.08%	26.15%	51.82%	63.17%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	27.52%	36.20%	58.27%	37.93%	48.35%	55.71%	13.59%	39.36%	51.27%	43.60%	62.99%	72.44%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	29.57%	37.97%	59.92%	40.08%	50.16%	54.41%	14.07%	41.01%	53.15%	45.49%	65.26%	74.40%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	21.96%	31.85%	52.65%	33.14%	45.58%	51.98%	12.61%	35.64%	47.50%	39.68%	59.62%	66.07%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	7.38%	16.65%	32.49%	15.40%	27.87%	45.90%	9.06%	24.42%	34.30%	24.96%	49.33%	58.64%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	29.47%	37.91%	59.85%	39.92%	50.13%	54.46%	14.05%	40.89%	52.98%	45.30%	65.04%	74.28%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	29.17%	37.66%	59.73%	39.49%	49.85%	54.06%	13.95%	40.63%	52.59%	44.95%	64.38%	73.87%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	28.77%	37.38%	59.44%	39.09%	49.74%	54.17%	13.89%	40.30%	52.20%	44.52%	64.06%	73.66%

tiled — 640×640 tiles, tiled weights

Size	Platform	Runtime	Decode	APs	Δ APs vs FP32 (pp)	AP	AP50	Time per image (ms)	Realized FPS	Source
Nano	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	19.46%	ref	28.10%	47.32%	36.17	55.5	v-1975
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	19.45%	-0.01	28.06%	47.27%	30.86	68.3	v-19cf
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	11.62%	-7.84	19.88%	39.80%	1623.36	0.9	v-1c40
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	11.62%	-7.84	19.88%	39.80%	514.63	2.1	v-1b53
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	11.62%	-7.84	19.88%	39.80%	515.42	2.1	v-1b5d
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	16.82%	-2.64	24.59%	43.26%	560.24	2.2	v-1b65
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	17.85%	-1.61	26.01%	45.71%	219.83	8.1	v-1773
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	17.85%	-1.61	26.01%	45.72%	253.65	8.9	v-1a1b
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	17.83%	-1.63	26.12%	45.93%	228.06	8.0	v-1aff
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	12.56%	-6.90	21.25%	42.33%	270.21	4.3	v-1af8
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	12.60%	-6.86	21.28%	42.34%	274.85	4.2	v-1af4
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	17.83%	-1.63	26.12%	45.93%	258.79	9.3	v-1b03
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	12.56%	-6.90	21.25%	42.33%	343.43	10.8	v-1afe
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	12.60%	-6.86	21.28%	42.35%	335.74	11.1	v-1af7
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	17.83%	-1.63	26.12%	45.93%	257.19	8.8	v-189f
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	17.85%	-1.61	26.01%	45.71%	230.52	8.1	v-178a
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	17.85%	-1.61	26.01%	45.71%	263.46	8.3	v-1a25
Nano	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	15.54%	-3.92	23.89%	44.19%	89.83	19.6	v-1692
Nano	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	15.54%	-3.92	23.89%	44.17%	73.38	49.0	v-1929
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	18.35%	-1.11	26.93%	46.58%	127.56	10.0	v-1633
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	19.42%	-0.04	28.06%	47.29%	132.93	44.1	v-1916
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	17.24%	-2.22	25.58%	45.11%	127.24	50.9	v-1918
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	13.57%	-5.89	22.68%	43.85%	41.14	105.8	v-1672
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	18.77%	-0.69	27.30%	46.70%	51.34	66.4	v-1673
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	19.45%	-0.01	28.02%	47.36%	95.37	45.9	v-1674
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	18.07%	-1.39	26.37%	46.20%	52.26	66.6	v-1d40
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	12.46%	-7.00	21.42%	42.72%	39.91	107.7	v-1d41
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	19.44%	-0.02	28.02%	47.36%	95.53	45.4	v-1d42
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	19.17%	-0.29	27.80%	47.14%	75.51	42.3	v-1d44
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	19.19%	-0.27	27.81%	47.20%	74.81	41.5	v-1d43
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	23.43%	ref	32.35%	53.28%	47.28	47.0	v-197b
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	23.37%	-0.06	32.31%	53.27%	36.33	61.0	v-19d3
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	12.50%	-10.93	21.01%	43.81%	2123.03	0.8	v-1c67
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	12.49%	-10.94	21.00%	43.81%	921.73	1.1	v-1bae
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	12.50%	-10.93	21.01%	43.85%	922.08	1.1	v-1bb7
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	18.50%	-4.93	26.17%	48.20%	956.87	1.2	v-1bc6
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	19.19%	-4.24	26.95%	49.51%	344.32	3.9	v-1787
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	19.19%	-4.24	26.94%	49.51%	608.99	6.5	v-1a21
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	19.53%	-3.90	27.15%	49.73%	358.04	3.7	v-1b44
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	13.18%	-10.25	21.97%	45.54%	422.97	2.6	v-1b3d
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	13.19%	-10.24	21.96%	45.56%	428.16	2.5	v-1b34
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	19.54%	-3.89	27.15%	49.73%	613.33	6.5	v-1b46
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	13.19%	-10.24	21.96%	45.54%	637.97	6.0	v-1b42
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	13.19%	-10.24	21.97%	45.56%	613.26	6.5	v-1b39
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	19.53%	-3.90	27.15%	49.73%	612.97	6.5	v-18a5
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	19.19%	-4.24	26.95%	49.51%	348.05	3.9	v-17be
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	19.20%	-4.23	26.95%	49.51%	610.06	6.4	v-1a2c
Small	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	18.04%	-5.39	25.91%	48.91%	102.83	16.3	v-161f
Small	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	18.07%	-5.36	25.91%	48.89%	81.87	27.7	v-194f
Small	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	21.71%	-1.72	30.13%	52.19%	233.92	4.9	v-163e
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	23.32%	-0.11	32.27%	53.18%	175.31	28.1	v-15d1
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	19.76%	-3.67	25.52%	47.28%	136.82	39.5	v-15d5
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	15.59%	-7.84	25.46%	48.84%	54.33	73.0	v-1698
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	21.89%	-1.54	30.55%	52.03%	63.06	70.9	v-1699
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	23.31%	-0.12	32.19%	53.09%	108.39	32.4	v-169a
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	20.14%	-3.29	28.13%	50.20%	62.38	72.4	v-1d83
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	13.91%	-9.52	22.60%	45.79%	48.17	74.4	v-1d84
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	23.29%	-0.14	32.18%	53.13%	106.87	32.5	v-1d85
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	23.05%	-0.38	32.06%	52.99%	81.11	30.1	v-1d88
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	23.16%	-0.27	32.15%	53.06%	84.11	30.2	v-1d87
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	26.74%	ref	35.08%	56.84%	96.88	32.5	v-1985
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	26.64%	-0.10	35.02%	56.81%	58.25	42.3	v-19d8
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	5.19%	-21.55	11.07%	24.14%	3036.26	0.6	v-1c80
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	5.19%	-21.55	11.05%	24.13%	1700.83	0.6	v-1c3f
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	5.19%	-21.55	11.06%	24.14%	1701.44	0.6	v-1c48
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	7.90%	-18.84 ⚠	14.05%	26.80%	1743.64	0.6	v-1c54
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	18.82%	-7.92	27.98%	51.93%	651.74	1.8	v-17ae
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	18.84%	-7.90	27.97%	51.93%	1772.40	2.3	v-1a31
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	19.26%	-7.48	28.20%	51.93%	664.30	1.7	v-1b96
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	12.03%	-14.71	21.32%	46.48%	731.55	1.4	v-1b8d
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	12.01%	-14.73	21.32%	46.45%	738.64	1.4	v-1b83
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	19.26%	-7.48	28.20%	51.93%	1753.30	2.3	v-1b9e
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	12.04%	-14.70	21.33%	46.48%	1777.57	2.2	v-1b92
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	12.01%	-14.73	21.32%	46.45%	1744.85	2.3	v-1b88
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	19.26%	-7.48	28.20%	51.93%	1751.32	2.3	v-18b7
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	18.84%	-7.90	27.97%	51.93%	651.93	1.9	v-17da
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	18.84%	-7.90	27.98%	51.93%	1753.64	2.3	v-1a38
Medium	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	13.21%	-13.53 ⚠	22.51%	47.86%	177.73	7.5	v-1631
Medium	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	13.12%	-13.62 ⚠	22.43%	47.80%	153.16	9.1	v-1954
Medium	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	23.43%	-3.31	30.57%	54.12%	420.05	2.6	v-1647
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	26.59%	-0.15	35.02%	56.77%	300.25	14.8	v-195a
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	17.26%	-9.48	24.14%	48.57%	213.91	22.3	v-195d
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	16.32%	-10.42	26.29%	51.46%	63.22	41.2	v-16dc
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	25.36%	-1.38	33.25%	55.87%	81.44	43.9	v-16e0
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	26.87%	+0.13	35.21%	56.91%	163.35	18.1	v-16e1
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	22.73%	-4.01	30.72%	54.77%	93.17	44.5	v-1dd5
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	15.03%	-11.71	23.82%	49.35%	61.37	41.9	v-1dd6
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	26.83%	+0.09	35.19%	56.89%	164.98	17.7	v-1dd7
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	26.46%	-0.28	34.99%	56.84%	132.75	17.4	v-1dd9
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	26.45%	-0.29	35.00%	56.80%	135.32	17.1	v-1dd8

Full metrics — tiled

Size	Platform	Runtime	Decode	Precision	APs	AP	AP50	AP75	APm	APl	AR@1	AR@10	AR@100	ARs	ARm	ARl
Nano	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	19.46%	28.10%	47.32%	28.15%	39.23%	46.33%	11.51%	33.47%	43.62%	34.90%	56.03%	57.43%
Nano	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	19.45%	28.06%	47.27%	28.20%	39.16%	46.23%	11.52%	33.41%	43.56%	34.86%	55.97%	57.28%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	11.62%	19.88%	39.80%	17.70%	30.67%	37.47%	9.33%	25.51%	32.61%	23.14%	45.67%	48.51%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	11.62%	19.88%	39.80%	17.70%	30.67%	37.47%	9.33%	25.51%	32.61%	23.14%	45.67%	48.51%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	11.62%	19.88%	39.80%	17.70%	30.66%	37.47%	9.33%	25.51%	32.61%	23.15%	45.66%	48.51%
Nano	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	16.82%	24.59%	43.26%	24.23%	34.42%	38.95%	10.53%	30.35%	39.41%	31.20%	50.36%	50.30%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	17.85%	26.01%	45.71%	25.83%	36.56%	44.28%	10.82%	31.45%	40.89%	32.45%	53.08%	53.64%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	17.85%	26.01%	45.72%	25.83%	36.57%	44.28%	10.82%	31.45%	40.89%	32.45%	53.08%	53.64%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	17.83%	26.12%	45.93%	25.86%	36.54%	45.35%	10.89%	31.59%	40.86%	32.49%	52.68%	55.91%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	12.56%	21.25%	42.33%	19.10%	32.82%	41.20%	9.46%	26.54%	34.00%	24.35%	47.89%	50.97%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	12.60%	21.28%	42.34%	19.14%	32.84%	41.19%	9.46%	26.54%	34.02%	24.39%	47.91%	50.97%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	17.83%	26.12%	45.93%	25.86%	36.54%	45.35%	10.89%	31.59%	40.86%	32.49%	52.68%	55.91%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	12.56%	21.25%	42.33%	19.10%	32.82%	41.18%	9.46%	26.54%	34.00%	24.35%	47.89%	50.97%
Nano	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	12.60%	21.28%	42.35%	19.13%	32.84%	41.19%	9.46%	26.54%	34.02%	24.39%	47.91%	50.97%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	17.83%	26.12%	45.93%	25.86%	36.54%	45.34%	10.89%	31.59%	40.86%	32.49%	52.68%	55.91%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	17.85%	26.01%	45.71%	25.83%	36.56%	44.28%	10.82%	31.45%	40.89%	32.45%	53.08%	53.64%
Nano	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	17.85%	26.01%	45.71%	25.83%	36.57%	44.28%	10.82%	31.45%	40.89%	32.45%	53.08%	53.64%
Nano	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	15.54%	23.89%	44.19%	22.88%	35.05%	40.90%	10.29%	29.20%	37.63%	28.89%	50.46%	52.37%
Nano	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	15.54%	23.89%	44.17%	22.84%	35.02%	41.10%	10.32%	29.20%	37.62%	28.89%	50.43%	52.39%
Nano	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	18.35%	26.93%	46.58%	26.93%	37.95%	46.03%	11.14%	32.07%	41.52%	33.02%	53.43%	57.90%
Nano	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	19.42%	28.06%	47.29%	28.21%	39.15%	46.20%	11.44%	33.45%	43.50%	34.75%	55.95%	57.14%
Nano	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	17.24%	25.58%	45.11%	25.13%	36.34%	45.02%	10.69%	31.19%	40.52%	32.13%	52.16%	54.99%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	13.57%	22.68%	43.85%	21.01%	34.81%	43.78%	9.95%	28.05%	35.82%	25.88%	50.05%	55.13%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	18.77%	27.30%	46.70%	27.52%	38.21%	45.17%	11.22%	32.76%	42.59%	34.01%	54.82%	56.54%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	19.45%	28.02%	47.36%	28.16%	39.16%	45.20%	11.46%	33.33%	43.53%	34.93%	55.77%	56.07%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	18.07%	26.37%	46.20%	26.50%	36.87%	46.44%	10.90%	31.87%	41.43%	33.00%	53.47%	56.52%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	12.46%	21.42%	42.72%	19.09%	33.23%	44.04%	9.56%	26.84%	34.30%	24.21%	48.90%	55.01%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	19.44%	28.02%	47.36%	28.16%	39.13%	45.84%	11.49%	33.36%	43.47%	34.78%	55.81%	57.30%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	19.17%	27.80%	47.14%	27.87%	39.05%	44.64%	11.39%	33.03%	43.04%	34.48%	55.12%	55.52%
Nano	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	19.19%	27.81%	47.20%	27.78%	39.04%	44.85%	11.43%	33.12%	43.05%	34.42%	55.23%	55.98%
Small	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	23.43%	32.35%	53.28%	33.13%	44.08%	47.11%	12.69%	36.75%	47.48%	39.35%	59.40%	63.35%
Small	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	23.37%	32.31%	53.27%	33.01%	44.09%	47.08%	12.69%	36.70%	47.44%	39.28%	59.39%	63.30%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	12.50%	21.01%	43.81%	18.09%	32.64%	35.16%	9.64%	27.34%	34.81%	25.58%	48.43%	47.50%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	12.49%	21.00%	43.81%	18.08%	32.64%	35.16%	9.64%	27.34%	34.81%	25.58%	48.43%	47.50%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	12.50%	21.01%	43.85%	18.09%	32.65%	35.16%	9.65%	27.35%	34.82%	25.58%	48.44%	47.50%
Small	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	18.50%	26.17%	48.20%	25.23%	36.55%	38.84%	11.01%	32.26%	41.61%	33.77%	52.90%	53.48%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	19.19%	26.95%	49.51%	26.35%	37.30%	34.89%	11.04%	32.71%	42.36%	34.83%	53.33%	49.32%
Small	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	19.19%	26.94%	49.51%	26.34%	37.31%	34.89%	11.04%	32.71%	42.36%	34.83%	53.33%	49.32%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	19.53%	27.15%	49.73%	26.58%	37.41%	36.05%	11.15%	33.00%	42.72%	35.23%	53.75%	51.18%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	13.18%	21.97%	45.54%	18.83%	34.31%	32.51%	9.85%	28.01%	35.74%	26.39%	49.75%	46.27%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	13.19%	21.96%	45.56%	18.80%	34.29%	32.65%	9.86%	28.01%	35.74%	26.39%	49.74%	46.41%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	19.54%	27.15%	49.73%	26.58%	37.40%	36.05%	11.15%	33.00%	42.72%	35.23%	53.75%	51.18%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	13.19%	21.96%	45.54%	18.83%	34.30%	32.51%	9.85%	28.01%	35.74%	26.39%	49.75%	46.27%
Small	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	13.19%	21.97%	45.56%	18.81%	34.30%	32.64%	9.86%	28.01%	35.74%	26.39%	49.74%	46.41%
Small	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	19.53%	27.15%	49.73%	26.58%	37.40%	36.05%	11.15%	33.00%	42.72%	35.23%	53.75%	51.18%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	19.19%	26.95%	49.51%	26.35%	37.30%	34.89%	11.04%	32.71%	42.36%	34.83%	53.33%	49.32%
Small	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	19.20%	26.95%	49.51%	26.35%	37.30%	34.86%	11.04%	32.71%	42.36%	34.83%	53.33%	49.32%
Small	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	18.04%	25.91%	48.91%	24.99%	37.24%	36.59%	10.90%	31.86%	41.00%	33.04%	53.09%	50.22%
Small	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	18.07%	25.91%	48.89%	25.01%	37.25%	36.34%	10.90%	31.85%	40.98%	33.07%	53.07%	49.91%
Small	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	21.71%	30.13%	52.19%	30.09%	41.37%	44.13%	12.00%	34.79%	44.71%	36.89%	56.18%	58.49%
Small	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	23.32%	32.27%	53.18%	32.96%	44.05%	46.92%	12.60%	36.74%	47.39%	39.18%	59.30%	63.42%
Small	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	19.76%	25.52%	47.28%	24.43%	35.81%	41.77%	10.70%	32.38%	42.30%	35.37%	52.80%	51.18%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	15.59%	25.46%	48.84%	23.40%	39.10%	44.50%	10.86%	30.32%	38.40%	28.22%	53.15%	61.86%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	21.89%	30.55%	52.03%	30.62%	42.26%	45.89%	12.19%	35.57%	45.86%	37.53%	58.07%	61.64%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	23.31%	32.19%	53.09%	33.09%	43.99%	45.03%	12.65%	36.70%	47.41%	39.28%	59.34%	61.27%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	20.14%	28.13%	50.20%	27.94%	38.85%	41.11%	11.48%	33.84%	43.74%	35.82%	55.44%	56.65%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	13.91%	22.60%	45.79%	20.43%	34.79%	36.00%	10.14%	28.82%	36.55%	27.20%	50.73%	50.20%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	23.29%	32.18%	53.13%	33.01%	44.06%	44.67%	12.64%	36.67%	47.38%	39.18%	59.38%	61.38%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	23.05%	32.06%	52.99%	32.88%	43.87%	45.29%	12.62%	36.46%	47.07%	38.95%	58.70%	61.32%
Small	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	23.16%	32.15%	53.06%	32.87%	44.02%	45.03%	12.61%	36.54%	47.23%	39.09%	59.01%	61.43%
Medium	ONNX FP32 (CUDA · RTX 3090)	ONNX	—	FP32	26.74%	35.08%	56.84%	36.47%	46.78%	52.16%	13.34%	38.99%	50.12%	42.90%	61.14%	66.89%
Medium	ONNX FP16 (CUDA · RTX 3090)	ONNX	—	FP16	26.64%	35.02%	56.81%	36.32%	46.75%	52.14%	13.32%	38.93%	50.06%	42.81%	61.08%	66.84%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (Ezurio) — latency	LiteRT	Logical	INT8	5.19%	11.07%	24.14%	9.06%	18.59%	19.87%	6.93%	18.82%	24.22%	15.01%	34.76%	29.39%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Combined	INT8	5.19%	11.05%	24.13%	9.02%	18.58%	19.87%	6.93%	18.82%	24.22%	15.01%	34.76%	29.39%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Logical	INT8	5.19%	11.06%	24.14%	9.05%	18.59%	19.87%	6.93%	18.82%	24.22%	15.01%	34.76%	29.39%
Medium	NXP i.MX 8M Plus + VeriSilicon NPU (FRDM)	LiteRT	Smart	INT8	7.90%	14.05%	26.80%	13.02%	21.79%	20.86%	7.84%	22.56%	29.54%	20.44%	40.09%	30.78%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — latency	LiteRT	Smart	INT8	18.82%	27.98%	51.93%	26.53%	40.82%	45.93%	11.54%	33.25%	42.36%	33.96%	55.50%	54.79%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Ezurio) — throughput	LiteRT	Smart	INT8	18.84%	27.97%	51.93%	26.53%	40.80%	45.93%	11.54%	33.25%	42.36%	33.96%	55.50%	54.79%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Smart	INT8	19.26%	28.20%	51.93%	27.18%	40.93%	44.34%	11.77%	33.41%	42.52%	34.17%	55.58%	56.00%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Logical	INT8	12.03%	21.32%	46.48%	17.15%	34.28%	42.50%	9.84%	26.93%	33.74%	24.53%	47.74%	55.23%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — latency	LiteRT	Combined	INT8	12.01%	21.32%	46.45%	17.14%	34.31%	42.50%	9.85%	26.93%	33.73%	24.52%	47.72%	55.23%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Smart	INT8	19.26%	28.20%	51.93%	27.18%	40.92%	44.34%	11.77%	33.41%	42.52%	34.17%	55.58%	56.00%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Logical	INT8	12.04%	21.33%	46.48%	17.16%	34.30%	42.49%	9.85%	26.93%	33.74%	24.53%	47.74%	55.23%
Medium	NXP i.MX 95 + eIQ Neutron NPU (EVK) — throughput	LiteRT	Combined	INT8	12.01%	21.32%	46.45%	17.14%	34.31%	42.50%	9.85%	26.93%	33.73%	24.52%	47.72%	55.23%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Phytec) — throughput	LiteRT	Smart	INT8	19.26%	28.20%	51.93%	27.18%	40.92%	44.34%	11.77%	33.41%	42.52%	34.17%	55.58%	56.00%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — latency	LiteRT	Smart	INT8	18.84%	27.97%	51.93%	26.53%	40.81%	45.93%	11.54%	33.25%	42.36%	33.96%	55.50%	54.79%
Medium	NXP i.MX 95 + eIQ Neutron NPU (Verdin) — throughput	LiteRT	Smart	INT8	18.84%	27.98%	51.93%	26.54%	40.82%	45.93%	11.54%	33.25%	42.36%	33.96%	55.50%	54.79%
Medium	NXP Ara240 (FRDM) — latency	Ara DVM	Smart	INT16	13.21%	22.51%	47.86%	18.47%	35.41%	43.36%	9.95%	28.14%	35.74%	26.77%	49.46%	55.77%
Medium	NXP Ara240 (FRDM) — throughput	Ara DVM	Smart	INT16	13.12%	22.43%	47.80%	18.43%	35.39%	43.30%	9.94%	28.12%	35.70%	26.67%	49.49%	55.74%
Medium	Raspberry Pi 5 + Hailo-8L NPU	Hailo HEF	—	INT8	23.43%	30.57%	54.12%	30.36%	41.61%	46.13%	11.99%	35.47%	45.46%	38.21%	56.86%	59.30%
Medium	NVIDIA Jetson Orin Nano (TensorRT FP16)	TensorRT	—	FP16	26.59%	35.02%	56.77%	36.38%	46.76%	52.65%	13.35%	38.87%	50.02%	42.63%	61.13%	66.42%
Medium	NVIDIA Jetson Orin Nano (TensorRT INT8)	TensorRT	—	INT8	17.26%	24.14%	48.57%	21.27%	35.84%	41.25%	10.06%	30.52%	39.41%	32.44%	49.91%	56.45%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — throughput preset	ONNX · QNN EP	Logical	INT8	16.32%	26.29%	51.46%	23.81%	40.16%	48.21%	11.26%	30.93%	38.88%	29.11%	53.77%	61.11%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · TFE + SeqMSE) — balanced preset	ONNX · QNN EP	Smart	INT8	25.36%	33.25%	55.87%	33.97%	44.57%	47.78%	12.97%	37.78%	48.61%	41.27%	60.01%	60.32%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard) — accuracy preset	ONNX · QNN EP	Smart	INT16	26.87%	35.21%	56.91%	36.51%	46.95%	49.49%	13.34%	39.12%	50.47%	43.29%	61.44%	65.93%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Smart	INT8	22.73%	30.72%	54.77%	30.39%	42.14%	46.52%	12.04%	35.05%	44.84%	37.21%	56.78%	57.99%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT8 · Standard)	ONNX · QNN EP	Logical	INT8	15.03%	23.82%	49.35%	20.17%	36.59%	44.35%	10.17%	29.08%	36.72%	27.67%	51.00%	55.33%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Smart	INT16	26.83%	35.19%	56.89%	36.64%	46.94%	48.68%	13.35%	39.07%	50.41%	43.17%	61.45%	64.94%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · Standard)	ONNX · QNN EP	Logical	INT16	26.46%	34.99%	56.84%	36.25%	46.73%	49.46%	13.30%	38.77%	49.99%	42.77%	60.81%	65.57%
Medium	Qualcomm Dragonwing IQ-9075 — Hexagon NPU (INT16 · TFE + SeqMSE)	ONNX · QNN EP	Logical	INT16	26.45%	35.00%	56.80%	36.33%	46.79%	48.87%	13.31%	38.81%	49.98%	42.69%	60.88%	64.60%

---

Validation pipeline

1. EdgeFirst Profiler runs on the target hardware, executes the full inference pipeline (capture → preprocess → inference → postprocess, with slicing and merging for the tiled variant), and scores the predictions against ground truth on the device.
2. EdgeFirst Studio receives the metrics, predictions, charts and trace, and publishes them as the v-XXXX validation session every row above cites.
3. EdgeFirst HAL (open source) provides the hardware-accelerated preprocessing and post-decoding primitives used at both validation and deployment time.

Validation runs decode a JPEG per image and evaluate at a 0.001 confidence threshold to capture every detection for mAP. A camera pipeline at a deployment threshold of 0.25–0.75 carries a lighter host and postprocess load.

---

Downloads

Every file below is the artifact a validation session in the tables above measured. Files are organized as {variant}/{target folder}/ and named visdrone-yolo11{size}-det-{precision}[-smart|-logical]-{variant}{extension}. Each model file embeds the EdgeFirst edgefirst.json metadata (training session, dataset version, calibration artifact, converter chain), so a single file is sufficient for deployment.

640/

ONNX FP32 — Any platform with ONNX Runtime.

Size	File	Variant	Download
Nano	640/onnx/visdrone-yolo11n-det-fp16-640.onnx	FP16	Download
Nano	640/onnx/visdrone-yolo11n-det-fp32-640.onnx	FP32	Download
Small	640/onnx/visdrone-yolo11s-det-fp16-640.onnx	FP16	Download
Small	640/onnx/visdrone-yolo11s-det-fp32-640.onnx	FP32	Download
Medium	640/onnx/visdrone-yolo11m-det-fp16-640.onnx	FP16	Download
Medium	640/onnx/visdrone-yolo11m-det-fp32-640.onnx	FP32	Download

TFLite INT8 — CPU or NPU via runtime delegate (NXP i.MX 8M Plus VX Delegate).

Size	File	Variant	Download
Nano	640/tflite/visdrone-yolo11n-det-int8-640.tflite	Logical · INT8	Download
Nano	640/tflite/visdrone-yolo11n-det-int8-smart-640.tflite	Smart · INT8	Download
Small	640/tflite/visdrone-yolo11s-det-int8-640.tflite	Logical · INT8	Download
Small	640/tflite/visdrone-yolo11s-det-int8-smart-640.tflite	Smart · INT8	Download
Medium	640/tflite/visdrone-yolo11m-det-int8-640.tflite	Logical · INT8	Download
Medium	640/tflite/visdrone-yolo11m-det-int8-smart-640.tflite	Smart · INT8	Download

NXP i.MX 95 (eIQ Neutron) — eIQ Neutron NPU optimized.

Size	File	Variant	Download
Nano	640/imx95/visdrone-yolo11n-det-int8-640.imx95.tflite	Logical · INT8	Download
Nano	640/imx95/visdrone-yolo11n-det-int8-smart-640.imx95.tflite	Smart · INT8	Download
Small	640/imx95/visdrone-yolo11s-det-int8-640.imx95.tflite	Logical · INT8	Download
Small	640/imx95/visdrone-yolo11s-det-int8-smart-640.imx95.tflite	Smart · INT8	Download
Medium	640/imx95/visdrone-yolo11m-det-int8-640.imx95.tflite	Logical · INT8	Download
Medium	640/imx95/visdrone-yolo11m-det-int8-smart-640.imx95.tflite	Smart · INT8	Download

NXP Ara240 — Ara240 DNPU compiled model.

Size	File	Variant	Download
Nano	640/ara240/visdrone-yolo11n-det-int16-640.dvm	Smart · INT16	Download
Small	640/ara240/visdrone-yolo11s-det-int16-640.dvm	Smart · INT16	Download
Medium	640/ara240/visdrone-yolo11m-det-int16-640.dvm	Smart · INT16	Download

RPi5 + Hailo-8L (13 TOPS) — Hailo-8L 13 TOPS accelerator.

Size	File	Variant	Download
Nano	640/hailo/visdrone-yolo11n-det-int8-640.hailo8l.hef	INT8	Download
Small	640/hailo/visdrone-yolo11s-det-int8-640.hailo8l.hef	INT8	Download
Medium	640/hailo/visdrone-yolo11m-det-int8-640.hailo8l.hef	INT8	Download

NVIDIA Jetson (TensorRT) — Jetson FP16 and INT8 engines.

Size	File	Variant	Download
Nano	640/jetson/visdrone-yolo11n-det-fp16-640.engine	FP16	Download
Nano	640/jetson/visdrone-yolo11n-det-int8-640.engine	INT8	Download
Small	640/jetson/visdrone-yolo11s-det-fp16-640.engine	FP16	Download
Small	640/jetson/visdrone-yolo11s-det-int8-640.engine	INT8	Download
Medium	640/jetson/visdrone-yolo11m-det-fp16-640.engine	FP16	Download
Medium	640/jetson/visdrone-yolo11m-det-int8-640.engine	INT8	Download

Qualcomm Dragonwing IQ-9075 (Hexagon v73) — Offline-compiled Hexagon context binary, run through the ONNX Runtime QNN execution provider. Loads only on the Dragonwing IQ-9075 (Hexagon v73).

Size	File	Variant	Download
Nano	640/qnn/visdrone-yolo11n-det-w8a16-logical-in8-640.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Nano	640/qnn/visdrone-yolo11n-det-w8a16-logical-tfe-seqmse-in8-640.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Nano	640/qnn/visdrone-yolo11n-det-w8a16-smart-in8-640.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Nano	640/qnn/visdrone-yolo11n-det-w8a16-smart-tfe-seqmse-in8-640.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Nano	640/qnn/visdrone-yolo11n-det-w8a8-logical-640.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Nano	640/qnn/visdrone-yolo11n-det-w8a8-logical-tfe-seqmse-640.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Nano	640/qnn/visdrone-yolo11n-det-w8a8-smart-640.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Nano	640/qnn/visdrone-yolo11n-det-w8a8-smart-tfe-seqmse-640.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Small	640/qnn/visdrone-yolo11s-det-w8a16-logical-in8-640.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Small	640/qnn/visdrone-yolo11s-det-w8a16-logical-tfe-seqmse-in8-640.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Small	640/qnn/visdrone-yolo11s-det-w8a16-smart-in8-640.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Small	640/qnn/visdrone-yolo11s-det-w8a16-smart-tfe-seqmse-in8-640.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Small	640/qnn/visdrone-yolo11s-det-w8a8-logical-640.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Small	640/qnn/visdrone-yolo11s-det-w8a8-logical-tfe-seqmse-640.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Small	640/qnn/visdrone-yolo11s-det-w8a8-smart-640.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Small	640/qnn/visdrone-yolo11s-det-w8a8-smart-tfe-seqmse-640.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Medium	640/qnn/visdrone-yolo11m-det-w8a16-logical-in8-640.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Medium	640/qnn/visdrone-yolo11m-det-w8a16-logical-tfe-seqmse-in8-640.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Medium	640/qnn/visdrone-yolo11m-det-w8a16-smart-in8-640.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Medium	640/qnn/visdrone-yolo11m-det-w8a16-smart-tfe-seqmse-in8-640.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Medium	640/qnn/visdrone-yolo11m-det-w8a8-logical-640.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Medium	640/qnn/visdrone-yolo11m-det-w8a8-logical-tfe-seqmse-640.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Medium	640/qnn/visdrone-yolo11m-det-w8a8-smart-640.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Medium	640/qnn/visdrone-yolo11m-det-w8a8-smart-tfe-seqmse-640.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download

1280/

ONNX FP32 — Any platform with ONNX Runtime.

Size	File	Variant	Download
Nano	1280/onnx/visdrone-yolo11n-det-fp16-1280.onnx	FP16	Download
Nano	1280/onnx/visdrone-yolo11n-det-fp32-1280.onnx	FP32	Download
Small	1280/onnx/visdrone-yolo11s-det-fp16-1280.onnx	FP16	Download
Small	1280/onnx/visdrone-yolo11s-det-fp32-1280.onnx	FP32	Download
Medium	1280/onnx/visdrone-yolo11m-det-fp16-1280.onnx	FP16	Download
Medium	1280/onnx/visdrone-yolo11m-det-fp32-1280.onnx	FP32	Download

TFLite INT8 — CPU or NPU via runtime delegate (NXP i.MX 8M Plus VX Delegate).

Size	File	Variant	Download
Nano	1280/tflite/visdrone-yolo11n-det-int8-1280.tflite	Logical · INT8	Download
Nano	1280/tflite/visdrone-yolo11n-det-int8-smart-1280.tflite	Smart · INT8	Download
Small	1280/tflite/visdrone-yolo11s-det-int8-1280.tflite	Logical · INT8	Download
Small	1280/tflite/visdrone-yolo11s-det-int8-smart-1280.tflite	Smart · INT8	Download
Medium	1280/tflite/visdrone-yolo11m-det-int8-1280.tflite	Logical · INT8	Download
Medium	1280/tflite/visdrone-yolo11m-det-int8-smart-1280.tflite	Smart · INT8	Download

NXP i.MX 95 (eIQ Neutron) — eIQ Neutron NPU optimized.

Size	File	Variant	Download
Nano	1280/imx95/visdrone-yolo11n-det-int8-1280.imx95.tflite	Logical · INT8	Download
Nano	1280/imx95/visdrone-yolo11n-det-int8-smart-1280.imx95.tflite	Smart · INT8	Download
Small	1280/imx95/visdrone-yolo11s-det-int8-1280.imx95.tflite	Logical · INT8	Download
Small	1280/imx95/visdrone-yolo11s-det-int8-smart-1280.imx95.tflite	Smart · INT8	Download
Medium	1280/imx95/visdrone-yolo11m-det-int8-1280.imx95.tflite	Logical · INT8	Download
Medium	1280/imx95/visdrone-yolo11m-det-int8-smart-1280.imx95.tflite	Smart · INT8	Download

RPi5 + Hailo-8L (13 TOPS) — Hailo-8L 13 TOPS accelerator.

Size	File	Variant	Download
Nano	1280/hailo/visdrone-yolo11n-det-int8-1280.hailo8l.hef	INT8	Download

NVIDIA Jetson (TensorRT) — Jetson FP16 and INT8 engines.

Size	File	Variant	Download
Nano	1280/jetson/visdrone-yolo11n-det-fp16-1280.engine	FP16	Download
Nano	1280/jetson/visdrone-yolo11n-det-int8-1280.engine	INT8	Download
Small	1280/jetson/visdrone-yolo11s-det-fp16-1280.engine	FP16	Download
Small	1280/jetson/visdrone-yolo11s-det-int8-1280.engine	INT8	Download
Medium	1280/jetson/visdrone-yolo11m-det-fp16-1280.engine	FP16	Download
Medium	1280/jetson/visdrone-yolo11m-det-int8-1280.engine	INT8	Download

Qualcomm Dragonwing IQ-9075 (Hexagon v73) — Offline-compiled Hexagon context binary, run through the ONNX Runtime QNN execution provider. Loads only on the Dragonwing IQ-9075 (Hexagon v73).

Size	File	Variant	Download
Nano	1280/qnn/visdrone-yolo11n-det-w8a16-logical-in8-1280.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Nano	1280/qnn/visdrone-yolo11n-det-w8a16-logical-tfe-seqmse-in8-1280.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Nano	1280/qnn/visdrone-yolo11n-det-w8a16-smart-in8-1280.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Nano	1280/qnn/visdrone-yolo11n-det-w8a16-smart-tfe-seqmse-in8-1280.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Nano	1280/qnn/visdrone-yolo11n-det-w8a8-logical-1280.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Nano	1280/qnn/visdrone-yolo11n-det-w8a8-logical-tfe-seqmse-1280.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Nano	1280/qnn/visdrone-yolo11n-det-w8a8-smart-1280.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Nano	1280/qnn/visdrone-yolo11n-det-w8a8-smart-tfe-seqmse-1280.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Small	1280/qnn/visdrone-yolo11s-det-w8a16-logical-in8-1280.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Small	1280/qnn/visdrone-yolo11s-det-w8a16-logical-tfe-seqmse-in8-1280.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Small	1280/qnn/visdrone-yolo11s-det-w8a16-smart-in8-1280.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Small	1280/qnn/visdrone-yolo11s-det-w8a16-smart-tfe-seqmse-in8-1280.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Small	1280/qnn/visdrone-yolo11s-det-w8a8-logical-1280.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Small	1280/qnn/visdrone-yolo11s-det-w8a8-logical-tfe-seqmse-1280.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Small	1280/qnn/visdrone-yolo11s-det-w8a8-smart-1280.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Small	1280/qnn/visdrone-yolo11s-det-w8a8-smart-tfe-seqmse-1280.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Medium	1280/qnn/visdrone-yolo11m-det-w8a16-logical-in8-1280.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Medium	1280/qnn/visdrone-yolo11m-det-w8a16-logical-tfe-seqmse-in8-1280.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Medium	1280/qnn/visdrone-yolo11m-det-w8a16-smart-in8-1280.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Medium	1280/qnn/visdrone-yolo11m-det-w8a16-smart-tfe-seqmse-in8-1280.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Medium	1280/qnn/visdrone-yolo11m-det-w8a8-logical-1280.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Medium	1280/qnn/visdrone-yolo11m-det-w8a8-logical-tfe-seqmse-1280.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Medium	1280/qnn/visdrone-yolo11m-det-w8a8-smart-1280.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Medium	1280/qnn/visdrone-yolo11m-det-w8a8-smart-tfe-seqmse-1280.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download

1344x768/

ONNX FP32 — Any platform with ONNX Runtime.

Size	File	Variant	Download
Nano	1344x768/onnx/visdrone-yolo11n-det-fp16-1344x768.onnx	FP16	Download
Nano	1344x768/onnx/visdrone-yolo11n-det-fp32-1344x768.onnx	FP32	Download
Small	1344x768/onnx/visdrone-yolo11s-det-fp16-1344x768.onnx	FP16	Download
Small	1344x768/onnx/visdrone-yolo11s-det-fp32-1344x768.onnx	FP32	Download
Medium	1344x768/onnx/visdrone-yolo11m-det-fp16-1344x768.onnx	FP16	Download
Medium	1344x768/onnx/visdrone-yolo11m-det-fp32-1344x768.onnx	FP32	Download

TFLite INT8 — CPU or NPU via runtime delegate (NXP i.MX 8M Plus VX Delegate).

Size	File	Variant	Download
Nano	1344x768/tflite/visdrone-yolo11n-det-int8-1344x768.tflite	Logical · INT8	Download
Nano	1344x768/tflite/visdrone-yolo11n-det-int8-smart-1344x768.tflite	Smart · INT8	Download
Small	1344x768/tflite/visdrone-yolo11s-det-int8-1344x768.tflite	Logical · INT8	Download
Small	1344x768/tflite/visdrone-yolo11s-det-int8-smart-1344x768.tflite	Smart · INT8	Download
Medium	1344x768/tflite/visdrone-yolo11m-det-int8-1344x768.tflite	Logical · INT8	Download
Medium	1344x768/tflite/visdrone-yolo11m-det-int8-smart-1344x768.tflite	Smart · INT8	Download

NXP i.MX 95 (eIQ Neutron) — eIQ Neutron NPU optimized.

Size	File	Variant	Download
Nano	1344x768/imx95/visdrone-yolo11n-det-int8-1344x768.imx95.tflite	Logical · INT8	Download
Nano	1344x768/imx95/visdrone-yolo11n-det-int8-smart-1344x768.imx95.tflite	Smart · INT8	Download
Small	1344x768/imx95/visdrone-yolo11s-det-int8-1344x768.imx95.tflite	Logical · INT8	Download
Small	1344x768/imx95/visdrone-yolo11s-det-int8-smart-1344x768.imx95.tflite	Smart · INT8	Download
Medium	1344x768/imx95/visdrone-yolo11m-det-int8-1344x768.imx95.tflite	Logical · INT8	Download
Medium	1344x768/imx95/visdrone-yolo11m-det-int8-smart-1344x768.imx95.tflite	Smart · INT8	Download

NXP Ara240 — Ara240 DNPU compiled model.

Size	File	Variant	Download
Nano	1344x768/ara240/visdrone-yolo11n-det-int16-1344x768.dvm	Smart · INT16	Download
Small	1344x768/ara240/visdrone-yolo11s-det-int16-1344x768.dvm	Smart · INT16	Download
Medium	1344x768/ara240/visdrone-yolo11m-det-int16-1344x768.dvm	Smart · INT16	Download

RPi5 + Hailo-8L (13 TOPS) — Hailo-8L 13 TOPS accelerator.

Size	File	Variant	Download
Nano	1344x768/hailo/visdrone-yolo11n-det-int8-1344x768.hailo8l.hef	INT8	Download
Medium	1344x768/hailo/visdrone-yolo11m-det-int8-1344x768.hailo8l.hef	INT8	Download

NVIDIA Jetson (TensorRT) — Jetson FP16 and INT8 engines.

Size	File	Variant	Download
Nano	1344x768/jetson/visdrone-yolo11n-det-fp16-1344x768.engine	FP16	Download
Nano	1344x768/jetson/visdrone-yolo11n-det-int8-1344x768.engine	INT8	Download
Small	1344x768/jetson/visdrone-yolo11s-det-fp16-1344x768.engine	FP16	Download
Small	1344x768/jetson/visdrone-yolo11s-det-int8-1344x768.engine	INT8	Download
Medium	1344x768/jetson/visdrone-yolo11m-det-fp16-1344x768.engine	FP16	Download
Medium	1344x768/jetson/visdrone-yolo11m-det-int8-1344x768.engine	INT8	Download

Qualcomm Dragonwing IQ-9075 (Hexagon v73) — Offline-compiled Hexagon context binary, run through the ONNX Runtime QNN execution provider. Loads only on the Dragonwing IQ-9075 (Hexagon v73).

Size	File	Variant	Download
Nano	1344x768/qnn/visdrone-yolo11n-det-w8a16-logical-in8-1344x768.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Nano	1344x768/qnn/visdrone-yolo11n-det-w8a16-logical-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Nano	1344x768/qnn/visdrone-yolo11n-det-w8a16-smart-in8-1344x768.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Nano	1344x768/qnn/visdrone-yolo11n-det-w8a16-smart-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Nano	1344x768/qnn/visdrone-yolo11n-det-w8a8-logical-1344x768.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Nano	1344x768/qnn/visdrone-yolo11n-det-w8a8-logical-tfe-seqmse-1344x768.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Nano	1344x768/qnn/visdrone-yolo11n-det-w8a8-smart-1344x768.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Nano	1344x768/qnn/visdrone-yolo11n-det-w8a8-smart-tfe-seqmse-1344x768.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Small	1344x768/qnn/visdrone-yolo11s-det-w8a16-logical-in8-1344x768.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Small	1344x768/qnn/visdrone-yolo11s-det-w8a16-logical-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Small	1344x768/qnn/visdrone-yolo11s-det-w8a16-smart-in8-1344x768.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Small	1344x768/qnn/visdrone-yolo11s-det-w8a16-smart-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Small	1344x768/qnn/visdrone-yolo11s-det-w8a8-logical-1344x768.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Small	1344x768/qnn/visdrone-yolo11s-det-w8a8-logical-tfe-seqmse-1344x768.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Small	1344x768/qnn/visdrone-yolo11s-det-w8a8-smart-1344x768.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Small	1344x768/qnn/visdrone-yolo11s-det-w8a8-smart-tfe-seqmse-1344x768.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Medium	1344x768/qnn/visdrone-yolo11m-det-w8a16-logical-in8-1344x768.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Medium	1344x768/qnn/visdrone-yolo11m-det-w8a16-logical-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Medium	1344x768/qnn/visdrone-yolo11m-det-w8a16-smart-in8-1344x768.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Medium	1344x768/qnn/visdrone-yolo11m-det-w8a16-smart-tfe-seqmse-in8-1344x768.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Medium	1344x768/qnn/visdrone-yolo11m-det-w8a8-logical-1344x768.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Medium	1344x768/qnn/visdrone-yolo11m-det-w8a8-logical-tfe-seqmse-1344x768.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Medium	1344x768/qnn/visdrone-yolo11m-det-w8a8-smart-1344x768.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Medium	1344x768/qnn/visdrone-yolo11m-det-w8a8-smart-tfe-seqmse-1344x768.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download

tiled/

ONNX FP32 — Any platform with ONNX Runtime.

Size	File	Variant	Download
Nano	tiled/onnx/visdrone-yolo11n-det-fp16-tiled.onnx	FP16	Download
Nano	tiled/onnx/visdrone-yolo11n-det-fp32-tiled.onnx	FP32	Download
Small	tiled/onnx/visdrone-yolo11s-det-fp16-tiled.onnx	FP16	Download
Small	tiled/onnx/visdrone-yolo11s-det-fp32-tiled.onnx	FP32	Download
Medium	tiled/onnx/visdrone-yolo11m-det-fp16-tiled.onnx	FP16	Download
Medium	tiled/onnx/visdrone-yolo11m-det-fp32-tiled.onnx	FP32	Download

TFLite INT8 — CPU or NPU via runtime delegate (NXP i.MX 8M Plus VX Delegate).

Size	File	Variant	Download
Nano	tiled/tflite/visdrone-yolo11n-det-int8-smart-tiled.tflite	Smart · INT8	Download
Nano	tiled/tflite/visdrone-yolo11n-det-int8-tiled.tflite	Logical · INT8	Download
Small	tiled/tflite/visdrone-yolo11s-det-int8-smart-tiled.tflite	Smart · INT8	Download
Small	tiled/tflite/visdrone-yolo11s-det-int8-tiled.tflite	Logical · INT8	Download
Medium	tiled/tflite/visdrone-yolo11m-det-int8-smart-tiled.tflite	Smart · INT8	Download
Medium	tiled/tflite/visdrone-yolo11m-det-int8-tiled.tflite	Logical · INT8	Download

NXP i.MX 95 (eIQ Neutron) — eIQ Neutron NPU optimized.

Size	File	Variant	Download
Nano	tiled/imx95/visdrone-yolo11n-det-int8-smart-tiled.imx95.tflite	Smart · INT8	Download
Nano	tiled/imx95/visdrone-yolo11n-det-int8-tiled.imx95.tflite	Logical · INT8	Download
Small	tiled/imx95/visdrone-yolo11s-det-int8-smart-tiled.imx95.tflite	Smart · INT8	Download
Small	tiled/imx95/visdrone-yolo11s-det-int8-tiled.imx95.tflite	Logical · INT8	Download
Medium	tiled/imx95/visdrone-yolo11m-det-int8-smart-tiled.imx95.tflite	Smart · INT8	Download
Medium	tiled/imx95/visdrone-yolo11m-det-int8-tiled.imx95.tflite	Logical · INT8	Download

NXP Ara240 — Ara240 DNPU compiled model.

Size	File	Variant	Download
Nano	tiled/ara240/visdrone-yolo11n-det-int16-tiled.dvm	Smart · INT16	Download
Small	tiled/ara240/visdrone-yolo11s-det-int16-tiled.dvm	Smart · INT16	Download
Medium	tiled/ara240/visdrone-yolo11m-det-int16-tiled.dvm	Smart · INT16	Download

RPi5 + Hailo-8L (13 TOPS) — Hailo-8L 13 TOPS accelerator.

Size	File	Variant	Download
Nano	tiled/hailo/visdrone-yolo11n-det-int8-tiled.hailo8l.hef	INT8	Download
Small	tiled/hailo/visdrone-yolo11s-det-int8-tiled.hailo8l.hef	INT8	Download
Medium	tiled/hailo/visdrone-yolo11m-det-int8-tiled.hailo8l.hef	INT8	Download

NVIDIA Jetson (TensorRT) — Jetson FP16 and INT8 engines.

Size	File	Variant	Download
Nano	tiled/jetson/visdrone-yolo11n-det-fp16-tiled.engine	FP16	Download
Nano	tiled/jetson/visdrone-yolo11n-det-int8-tiled.engine	INT8	Download
Small	tiled/jetson/visdrone-yolo11s-det-fp16-tiled.engine	FP16	Download
Small	tiled/jetson/visdrone-yolo11s-det-int8-tiled.engine	INT8	Download
Medium	tiled/jetson/visdrone-yolo11m-det-fp16-tiled.engine	FP16	Download
Medium	tiled/jetson/visdrone-yolo11m-det-int8-tiled.engine	INT8	Download

Qualcomm Dragonwing IQ-9075 (Hexagon v73) — Offline-compiled Hexagon context binary, run through the ONNX Runtime QNN execution provider. Loads only on the Dragonwing IQ-9075 (Hexagon v73).

Size	File	Variant	Download
Nano	tiled/qnn/visdrone-yolo11n-det-w8a16-logical-in8-tiled.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Nano	tiled/qnn/visdrone-yolo11n-det-w8a16-logical-tfe-seqmse-in8-tiled.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Nano	tiled/qnn/visdrone-yolo11n-det-w8a16-smart-in8-tiled.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Nano	tiled/qnn/visdrone-yolo11n-det-w8a16-smart-tfe-seqmse-in8-tiled.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Nano	tiled/qnn/visdrone-yolo11n-det-w8a8-logical-tfe-seqmse-tiled.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Nano	tiled/qnn/visdrone-yolo11n-det-w8a8-logical-tiled.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Nano	tiled/qnn/visdrone-yolo11n-det-w8a8-smart-tfe-seqmse-tiled.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Nano	tiled/qnn/visdrone-yolo11n-det-w8a8-smart-tiled.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Small	tiled/qnn/visdrone-yolo11s-det-w8a16-logical-in8-tiled.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Small	tiled/qnn/visdrone-yolo11s-det-w8a16-logical-tfe-seqmse-in8-tiled.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Small	tiled/qnn/visdrone-yolo11s-det-w8a16-smart-in8-tiled.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Small	tiled/qnn/visdrone-yolo11s-det-w8a16-smart-tfe-seqmse-in8-tiled.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Small	tiled/qnn/visdrone-yolo11s-det-w8a8-logical-tfe-seqmse-tiled.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Small	tiled/qnn/visdrone-yolo11s-det-w8a8-logical-tiled.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Small	tiled/qnn/visdrone-yolo11s-det-w8a8-smart-tfe-seqmse-tiled.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Small	tiled/qnn/visdrone-yolo11s-det-w8a8-smart-tiled.iq9075.qnn.onnx	Smart · INT8 · Standard	Download
Medium	tiled/qnn/visdrone-yolo11m-det-w8a16-logical-in8-tiled.iq9075.qnn.onnx	Logical · INT16 · Standard	Download
Medium	tiled/qnn/visdrone-yolo11m-det-w8a16-logical-tfe-seqmse-in8-tiled.iq9075.qnn.onnx	Logical · INT16 · TFE + SeqMSE	Download
Medium	tiled/qnn/visdrone-yolo11m-det-w8a16-smart-in8-tiled.iq9075.qnn.onnx	Smart · INT16 · Standard	Download
Medium	tiled/qnn/visdrone-yolo11m-det-w8a16-smart-tfe-seqmse-in8-tiled.iq9075.qnn.onnx	Smart · INT16 · TFE + SeqMSE	Download
Medium	tiled/qnn/visdrone-yolo11m-det-w8a8-logical-tfe-seqmse-tiled.iq9075.qnn.onnx	Logical · INT8 · TFE + SeqMSE	Download
Medium	tiled/qnn/visdrone-yolo11m-det-w8a8-logical-tiled.iq9075.qnn.onnx	Logical · INT8 · Standard	Download
Medium	tiled/qnn/visdrone-yolo11m-det-w8a8-smart-tfe-seqmse-tiled.iq9075.qnn.onnx	Smart · INT8 · TFE + SeqMSE	Download
Medium	tiled/qnn/visdrone-yolo11m-det-w8a8-smart-tiled.iq9075.qnn.onnx	Smart · INT8 · Standard	Download

---

Traceability

Every measurement in the tables above is reachable through the EdgeFirst Studio validation framework. The v-XXXX Source link on each row resolves to a public Studio URL of the form:

https://edgefirst.studio/public/validation/v-XXXX/details?mode=charts

From there, the full provenance chain is one click deeper: training session ID, dataset version, calibration artifact, converter chain, validation parameters, and the host hardware description. The model file you download from this repository embeds the same chain in its edgefirst.json metadata.

---

See also

Model	Dataset	Link
YOLO11 Detection	COCO	EdgeFirst/yolo11-det
YOLOv8 Detection	VisDrone	EdgeFirst/visdrone-yolov8-det
YOLO26 Detection	VisDrone	EdgeFirst/visdrone-yolo26-det

---

Limitations

• Aerial bias — VisDrone is drone footage from Chinese cities: the scenes, object scales, viewpoints and the 10 classes do not transfer to ground-level cameras without retraining.
• Quantization loss — integer quantization costs accuracy against FP32, most visibly on small objects; the Δ APs vs FP32 column shows it per platform.
• Memory limits — the 1280 and 1344x768 variants exceed some accelerators' on-chip memory or attention limits; those cells read Not supported with the reason.

---

License

Model weights in this repository are derived from Ultralytics YOLO and remain © Ultralytics Inc., licensed AGPL-3.0 — use requires AGPL-3.0 compliance or an Ultralytics Enterprise License. The VisDrone dataset is © the AISKYEYE team, Tianjin University, and is subject to its own terms.

The validation results, this model card, and its metadata are Au-Zone Technologies' own contribution, licensed CC BY-NC 4.0 (Attribution — NonCommercial) — see the repository LICENSE for the full text and citation requirements.

---

Citation

@software{edgefirst_visdrone_yolo11_det,
  title = { {YOLO11 Detection on VisDrone — EdgeFirst Model Zoo} },
  author = {Au-Zone Technologies},
  url = {https://huggingface.co/EdgeFirst/visdrone-yolo11-det},
  year = {2026},
  license = {CC-BY-NC-4.0},
}

---

_{EdgeFirst Studio · GitHub · Docs · Au-Zone Technologies
Model weights © Ultralytics Inc. (AGPL-3.0) · Validation results & card © 2026 Au-Zone Technologies (CC BY-NC 4.0)
NXP^®, i.MX, eIQ^®, Neutron, and Ara240 are trademarks or products of NXP Semiconductors. Hailo is a trademark of Hailo Technologies Ltd. Jetson is a trademark of NVIDIA Corporation. Qualcomm and Dragonwing are trademarks of Qualcomm Incorporated. All other trademarks are the property of their respective owners.}