model_id,artifact_id,model_name,task,architecture_family,source_framework,source_repository,license,source_license,public_quantized_available,public_quantized_status,public_quantized_source,public_quantized_artifact,public_quantized_format,public_quantized_version,public_quantized_checksum,public_quantized_license,quantization_scheme,quantization_granularity,quantization_scope,weight_dtype,activation_dtype,input_dtype,output_dtype,integer_only_claimed,runtime_validation,runtime,runtime_version,paired_fp32_available,paired_fp32_source,paired_fp32_artifact,paired_fp32_format,paired_fp32_version,paired_fp32_checksum,paired_fp32_license,pair_compatibility,dataset,dataset_license,metric,input_shape,output_shape,preprocessing,postprocessing,label_space,dynamic_shape,custom_operator,control_flow,onnx_candidate,tflite_candidate,mlir_candidate,expected_ir_routes,artifact_local_path,paired_fp32_local_path,evidence_url,evidence_path,download_command,http_status,eligibility,priority,status_reason,notes
AD01,AD01-mlcommons-4addd0f-int8-tflite,MLPerf Tiny Deep AutoEncoder (DCASE 2020 ToyCar),anomaly_detection,Dense AutoEncoder / symmetric fully-connected 640-128-128-128-128-8-128-128-128-128-640; BatchNorm folded into Dense for TFLite,TensorFlow/Keras -> TensorFlow Lite,https://github.com/mlcommons/tiny/tree/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/anomaly_detection,MIT,MIT,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/anomaly_detection/trained_models/ad01_int8.tflite,research/downloads/anomaly_detection/AD01/ad01_int8.tflite,TFLite FlatBuffer,git:4addd0fa08d216e20637637874e084895f289da4; commit timestamp 2026-07-13T15:39:39Z,87cf24194ef93d1d9b11a591d805526b98008e351655d29883c825c9c106ba24,MIT task-local anomaly_detection license,FULL_INTEGER static PTQ performed upstream; published artifact consumed unchanged; TFLITE_BUILTINS_INT8 with int8 I/O,per-output-channel affine int8 FC weights (quantized_dimension=0 and scale_count equals output channels); per-tensor activations/I/O; int32 bias,"all 10 FC weights and biases, intermediate activations, input, and output; measured tensor dtypes are 21 int8 plus 10 int32",int8 weights; int32 biases,int8,int8; scale=0.3910152316093445; zero_point=89,int8; scale=0.36449846625328064; zero_point=96,TRUE; INDEPENDENTLY VERIFIED ALL GRAPH TENSORS INT8/INT32 AND ALL 10 BUILTINS FULLY_CONNECTED,"PASS: ai-edge-litert 2.1.6 load, allocate_tensors, semantic-zero input, and invoke for both public INT8 and paired FP32; all six local/fresh artifact copies match expected checksums",ai-edge-litert (LiteRT Interpreter),2.1.6,TRUE,https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/anomaly_detection/trained_models/ad01_fp32.tflite,research/downloads/anomaly_detection/AD01/ad01_fp32.tflite,TFLite FlatBuffer FP32,git:4addd0fa08d216e20637637874e084895f289da4; same trained_models release as INT8,c66636f4d7f8af8b10518e7be750a22c9d8d46ec97326b40b0d94c097e0aad9b,MIT task-local anomaly_detection license,VERIFIED,DCASE 2020 Task 2 ToyCar development dataset v1.0 (DOI 10.5281/zenodo.3678171) plus additional training dataset v1.0 (DOI 10.5281/zenodo.3727685),Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) for both cited Zenodo records,AUC and partial AUC from reconstruction error,"allocated [1,640], TFLite shape_signature [-1,640]","[1,640]",Upstream MLPerf Tiny DCASE pipeline: 640-element normalized log-mel feature vector,Task-specific postprocessing is defined by the corresponding pipeline config,640-element reconstructed feature vector; anomaly score is computed outside the network,FALSE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/anomaly_detection/AD01/ad01_int8.tflite,research/downloads/anomaly_detection/AD01/ad01_fp32.tflite,https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/anomaly_detection/trained_models/ad01_int8.tflite,research/evidence/anomaly_detection/independent_audit/ad01_independent_audit.json;research/evidence/anomaly_detection/independent_audit/execution_manifest.json;research/evidence/anomaly_detection/independent_audit/logs/checksums.stdout.log;research/evidence/anomaly_detection/independent_audit/logs/evidence_extract.stdout.log;research/evidence/anomaly_detection/independent_audit/sources/anomaly_detection_LICENSE;research/evidence/anomaly_detection/independent_audit/sources/baseline.yaml;research/evidence/anomaly_detection/independent_audit/sources/common.py;research/evidence/anomaly_detection/independent_audit/sources/keras_model.py;research/evidence/anomaly_detection/independent_audit/sources/02_convert.py,curl -fL 'https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/anomaly_detection/trained_models/ad01_int8.tflite',200,ELIGIBLE,P0,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
LM04,LM04-public-quantized-01,BERT-tiny RAID text detector ONNX,language_modeling/text_classification,encoder-only BERT Transformer; 2 layers; hidden 128; 2 heads; classification head,PyTorch Transformers -> ONNX,https://huggingface.co/ShantanuT01/BERT-tiny-RAID/tree/cbc7e63d4749e8a84f2f07d23576639ef979d6fa; https://huggingface.co/onnx-community/BERT-tiny-RAID-ONNX/tree/8f5741a3d45781899100c9a299b25624b5afa914,MIT,MIT,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://huggingface.co/onnx-community/BERT-tiny-RAID-ONNX/resolve/8f5741a3d45781899100c9a299b25624b5afa914/onnx/model_int8.onnx,research/downloads/language_model/LM04_model_int8.onnx,ONNX,8f5741a3d45781899100c9a299b25624b5afa914,1c06c10cc1b8c1be81e863e11c92fb65deb2f39deb3647284ebe590483266760,MIT (ONNX repository metadata and downloaded README),dynamic ONNX quantization; MatMulInteger/DynamicQuantizeLinear with quantized embeddings,per-tensor scalar weight scales observed for embeddings and MatMul/Gemm paths,14 MatMulInteger paths and 3 DequantizeLinear embedding paths; normalization/softmax/residual/classifier output remain FLOAT,INT8 MatMul weights; UINT8 embedding tables,dynamic UINT8 on selected MatMul paths plus FLOAT remainder,INT64 input_ids/attention_mask/token_type_ids,FLOAT single logit,FALSE,PASS: ONNX checker full_check; ONNX Runtime CPU load and two bitwise-identical invokes for both FP32/INT8,onnx 1.18.0; onnxruntime 1.22.1; numpy 2.2.6; CPUExecutionProvider; one thread,onnx 1.18.0; onnxruntime 1.22.1; numpy 2.2.6; CPUExecutionProvider; one thread,TRUE,https://huggingface.co/onnx-community/BERT-tiny-RAID-ONNX/resolve/8f5741a3d45781899100c9a299b25624b5afa914/onnx/model.onnx,research/downloads/language_model/LM04_model_fp32.onnx,ONNX FP32,8f5741a3d45781899100c9a299b25624b5afa914,b7b83d30c66fa5aef49fb3019ff294e451983c9a810e9999c8480c25af514d24,MIT,VERIFIED,RAID stratified training subset (upstream model-card claim),MIT (liamdugan/raid HF dataset metadata),"Single classification logit; fixed-input sign and absolute logit difference for pairing, RAID quality metric not evaluated","dynamic [batch,sequence] x3; max position embeddings 512; validation [1,6]","[""DYNAMIC"",1]","Pinned BERT uncased WordPiece vocabulary; [CLS] and [SEP]; INT64 input_ids, attention_mask, and token_type_ids",Task-specific postprocessing is defined by the corresponding pipeline config,One FLOAT logit; upstream model card describes likelihood that text is human,TRUE,UNKNOWN,FALSE,TRUE,UNKNOWN,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/language_model/LM04_model_int8.onnx,research/downloads/language_model/LM04_model_fp32.onnx,https://huggingface.co/onnx-community/BERT-tiny-RAID-ONNX/resolve/8f5741a3d45781899100c9a299b25624b5afa914/onnx/model_int8.onnx,research/evidence/language_model/sources/LM04_README.md; research/evidence/language_model/sources/hf_api_onnx-community_BERT-tiny-RAID.json; research/evidence/language_model/runtime/LM04_int8_runtime.json; research/evidence/language_model/runtime/LM04_fp32_runtime.json; research/evidence/language_model/runtime/LM04_pair_compatibility.json; research/evidence/language_model/runtime/LM04_quantization_inspection.json,curl -fL 'https://huggingface.co/onnx-community/BERT-tiny-RAID-ONNX/resolve/8f5741a3d45781899100c9a299b25624b5afa914/onnx/model_int8.onnx',200,ELIGIBLE,P0,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
OD06,OD06-public-quantized-01,EfficientDet-Lite0 int8/float32 v1,object_detection,EfficientNet-Lite0 depthwise CNN + BiFPN + shared detection heads,TensorFlow Lite / LiteRT,https://www.kaggle.com/models/tensorflow/efficientdet,Apache-2.0 (TensorFlow EfficientDet model card/model family); COCO images retain source-owner/Flickr terms,Apache-2.0 (TensorFlow EfficientDet model card/model family); COCO images retain source-owner/Flickr terms,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://developers.google.com/edge/mediapipe/solutions/vision/object_detector,research/downloads/detection/efficientdet_lite0_int8_v1.tflite,TFLite FlatBuffer with metadata,int8/1,0720bf247bd76e6594ea28fa9c6f7c5242be774818997dbbeffc4da460c723bb,Apache-2.0 (TensorFlow EfficientDet model card/model family); COCO images retain source-owner/Flickr terms,FULL_INTEGER core with dequantized float raw-head outputs,per-axis weights + per-tensor activations verified (324/589 quantized tensors have multiple scales),weights and internal activations; raw outputs dequantized to float32,int8 (plus int32 bias),int8,uint8; scale=0.0078125; zero_point=127,float32 class scores and box regressions,NO; OUTPUT BOUNDARY IS FLOAT32,PASS: int8 and FP32 ai-edge-litert 2.1.6 allocate+zero-input invoke; pair comparison exit_code=0,PASS: int8 and FP32 ai-edge-litert 2.1.6 allocate+zero-input invoke; pair comparison exit_code=0,PASS: int8 and FP32 ai-edge-litert 2.1.6 allocate+zero-input invoke; pair comparison exit_code=0,TRUE,https://storage.googleapis.com/mediapipe-models/object_detector/efficientdet_lite0/float32/1/efficientdet_lite0.tflite,research/downloads/detection/efficientdet_lite0_float32_v1.tflite,TFLite FlatBuffer,int8/1,40338edf5ec70d43e318b0a716a84d4564cd1802759a7a07170c7e43796dbf58,Apache-2.0 (TensorFlow EfficientDet model card/model family); COCO images retain source-owner/Flickr terms,VERIFIED,COCO 2017; 80 foreground labels represented by a 90-index embedded label map,UNKNOWN,COCO bbox mAP,"[1,320,320,3]","[[1,19206,90],[1,19206,4]]","RGB resize to 320x320; FP32 input float32, public quantized input uint8 with real=(q-127)/128; paired artifacts have identical spatial input and embedded metadata",Task-specific postprocessing is defined by the corresponding pipeline config,90-index embedded COCO label map representing 80 foreground categories; label bytes are identical in the FP32 and public-quantized artifacts,FALSE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/detection/efficientdet_lite0_int8_v1.tflite,research/downloads/detection/efficientdet_lite0_float32_v1.tflite,https://developers.google.com/edge/mediapipe/solutions/vision/object_detector,research/evidence/detection/sources/google_ai_edge_object_detector.html;research/evidence/detection/sources/efficientdet_lite0_model_card.html;research/evidence/detection/sources/kaggle_efficientdet_model_api.json;research/evidence/detection/logs/download_checksums.tsv;research/evidence/detection/runtime/OD06-public-quantized-01.json;research/evidence/detection/runtime/OD06-fp32-01.json;research/evidence/detection/runtime/pair_compatibility.json;research/evidence/detection/runtime/quantization_summary.json,curl -fL 'https://developers.google.com/edge/mediapipe/solutions/vision/object_detector',200,ELIGIBLE,P0,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
OD07,OD07-public-quantized-01,EfficientDet-Lite2 int8/float32 v1,object_detection,EfficientNet-Lite2 depthwise CNN + BiFPN + shared detection heads,TensorFlow Lite / LiteRT,https://www.kaggle.com/models/tensorflow/efficientdet,Apache-2.0 (TensorFlow EfficientDet model card/model family); COCO images retain source-owner/Flickr terms,Apache-2.0 (TensorFlow EfficientDet model card/model family); COCO images retain source-owner/Flickr terms,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://developers.google.com/edge/mediapipe/solutions/vision/object_detector,research/downloads/detection/efficientdet_lite2_int8_v1.tflite,TFLite FlatBuffer with metadata,int8/1,b3f50554cb0ea559e90328845f7d9ba4d13c8bff372914d24e06bc8bb72fa896,Apache-2.0 (TensorFlow EfficientDet model card/model family); COCO images retain source-owner/Flickr terms,FULL_INTEGER core with dequantized float raw-head outputs,per-axis weights + per-tensor activations verified (418/773 quantized tensors have multiple scales),weights and internal activations; raw outputs dequantized to float32,int8 (plus int32 bias),int8,uint8; scale=0.0078125; zero_point=127,float32 class scores and box regressions,NO; OUTPUT BOUNDARY IS FLOAT32,PASS: int8 and FP32 ai-edge-litert 2.1.6 allocate+zero-input invoke; pair comparison exit_code=0,PASS: int8 and FP32 ai-edge-litert 2.1.6 allocate+zero-input invoke; pair comparison exit_code=0,PASS: int8 and FP32 ai-edge-litert 2.1.6 allocate+zero-input invoke; pair comparison exit_code=0,TRUE,https://storage.googleapis.com/mediapipe-models/object_detector/efficientdet_lite2/float32/1/efficientdet_lite2.tflite,research/downloads/detection/efficientdet_lite2_float32_v1.tflite,TFLite FlatBuffer,int8/1,ad2abbf2b4e10585e15176fd7b5ef03c28dda959ae26fc142549fdd1814db91d,Apache-2.0 (TensorFlow EfficientDet model card/model family); COCO images retain source-owner/Flickr terms,VERIFIED,COCO 2017; 80 foreground labels represented by a 90-index embedded label map,UNKNOWN,COCO bbox mAP,"[1,448,448,3]","[[1,37629,90],[1,37629,4]]","RGB resize to 448x448; FP32 input float32, public quantized input uint8 with real=(q-127)/128; paired artifacts have identical spatial input and embedded metadata",Task-specific postprocessing is defined by the corresponding pipeline config,90-index embedded COCO label map representing 80 foreground categories; label bytes are identical in the FP32 and public-quantized artifacts,FALSE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/detection/efficientdet_lite2_int8_v1.tflite,research/downloads/detection/efficientdet_lite2_float32_v1.tflite,https://developers.google.com/edge/mediapipe/solutions/vision/object_detector,research/evidence/detection/sources/google_ai_edge_object_detector.html;research/evidence/detection/sources/efficientdet_lite2_model_card.html;research/evidence/detection/sources/kaggle_efficientdet_model_api.json;research/evidence/detection/logs/download_checksums.tsv;research/evidence/detection/runtime/OD07-public-quantized-01.json;research/evidence/detection/runtime/OD07-fp32-01.json;research/evidence/detection/runtime/pair_compatibility.json;research/evidence/detection/runtime/quantization_summary.json,curl -fL 'https://developers.google.com/edge/mediapipe/solutions/vision/object_detector',200,ELIGIBLE,P1,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
SG06,SG06-public-quantized-01,DeepLabV3-MobileNetV2 width 1.0 Pascal train_aug,semantic_segmentation,DeepLabV3-style MobileNetV2 depthwise CNN (no ASPP/decoder in this fast checkpoint),TensorFlow 1 frozen GraphDef / TFLite,https://github.com/tensorflow/models/tree/archive/research/deeplab,Apache-2.0 for TensorFlow/Coral model artifacts; PASCAL VOC images retain Flickr/source terms,Apache-2.0 for TensorFlow/Coral model artifacts; PASCAL VOC images retain Flickr/source terms,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://coral.ai/models/semantic-segmentation/,research/downloads/segmentation/deeplabv3_mnv2_pascal_quant.tflite,TFLite FlatBuffer,google-coral/test_data@104342d2d3480b3e66203073dac24f4e2dbb4c41; FP32 checkpoint 2018-01-29,350af47fd346ea6d74b58833a351f583ba860e9806ec5bd292ff8d8f1f249398,Apache-2.0 for TensorFlow/Coral model artifacts; PASCAL VOC images retain Flickr/source terms,FULL_INTEGER,per-tensor verified (182 quantized tensors; no multi-scale tensor),"weights, activations, and input; integer ArgMax output",uint8 (plus int32 bias),uint8,uint8; scale=0.0078125; zero_point=128,int64 category IDs,TRUE,PASS: quant ai-edge-litert 2.1.6 and FP32 tensorflow-cpu 2.21.0 inference; pair comparison exit_code=0,PASS: quant ai-edge-litert 2.1.6 and FP32 tensorflow-cpu 2.21.0 inference; pair comparison exit_code=0,PASS: quant ai-edge-litert 2.1.6 and FP32 tensorflow-cpu 2.21.0 inference; pair comparison exit_code=0,TRUE,https://storage.googleapis.com/download.tensorflow.org/models/deeplabv3_mnv2_pascal_train_aug_2018_01_29.tar.gz,research/downloads/segmentation/fp32_extracted/deeplabv3_mnv2_pascal_train_aug/frozen_inference_graph.pb,TensorFlow GraphDef/SavedModel,google-coral/test_data@104342d2d3480b3e66203073dac24f4e2dbb4c41; FP32 checkpoint 2018-01-29,b3b7c39d1010c6da1dd0e87973ca24a78fa7dc70d4136fbcf50b9585469b6d9a,Apache-2.0 for TensorFlow/Coral model artifacts; PASCAL VOC images retain Flickr/source terms,VERIFIED,"PASCAL VOC 2012 train_aug with ImageNet/MS-COCO pretraining; 21 IDs including background, ignore label 255",UNKNOWN,mean Intersection-over-Union,"quant [1,513,513,3]; FP32 [1,H,W,3] invoked at 513x513","[1,513,513]",RGB uint8; FP32 GraphDef has dynamic spatial input and was invoked at 513x513; public quantized artifact fixes 513x513 with scale=1/128 and zero_point=128,Task-specific postprocessing is defined by the corresponding pipeline config,21 PASCAL VOC category IDs including background; ignore label 255,FALSE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/segmentation/deeplabv3_mnv2_pascal_quant.tflite,research/downloads/segmentation/fp32_extracted/deeplabv3_mnv2_pascal_train_aug/frozen_inference_graph.pb,https://coral.ai/models/semantic-segmentation/,research/evidence/segmentation/sources/coral_semantic_segmentation.html;research/evidence/segmentation/sources/coral_test_data_LICENSE;research/evidence/segmentation/logs/download_checksums.tsv;research/evidence/segmentation/runtime/quantization_summary.json;research/evidence/segmentation/sources/tensorflow_deeplab_model_zoo.md;research/evidence/segmentation/sources/pascal_voc2012_dataset.html;research/evidence/segmentation/sources/pascal_voc_segmentation_labels.txt;research/evidence/segmentation/runtime/SG06-public-quantized-01.json;research/evidence/segmentation/runtime/SG06-fp32-01.json;research/evidence/segmentation/runtime/pair_compatibility.json,curl -fL 'https://coral.ai/models/semantic-segmentation/',200,ELIGIBLE,P0,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
SG07,SG07-public-quantized-01,DeepLabV3-MobileNetV2 width 0.5 Pascal train_aug,semantic_segmentation,DeepLabV3-style MobileNetV2 width-0.5 depthwise CNN (no ASPP/decoder in this fast checkpoint),TensorFlow 1 frozen GraphDef / TFLite,https://github.com/tensorflow/models/tree/archive/research/deeplab,Apache-2.0 for TensorFlow/Coral model artifacts; PASCAL VOC images retain Flickr/source terms,Apache-2.0 for TensorFlow/Coral model artifacts; PASCAL VOC images retain Flickr/source terms,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://coral.ai/models/semantic-segmentation/,research/downloads/segmentation/deeplabv3_mnv2_dm05_pascal_quant.tflite,TFLite FlatBuffer,google-coral/test_data@104342d2d3480b3e66203073dac24f4e2dbb4c41; FP32 checkpoint 2018-10-01,0470d2a782aa54eeeb99d32e7b6b3fb7722905c7c4f5d26bd957ec861366d48b,Apache-2.0 for TensorFlow/Coral model artifacts; PASCAL VOC images retain Flickr/source terms,FULL_INTEGER,per-tensor verified (182 quantized tensors; no multi-scale tensor),"weights, activations, and input; integer ArgMax output",uint8 (plus int32 bias),uint8,uint8; scale=0.0078125; zero_point=128,int64 category IDs,TRUE,PASS: quant ai-edge-litert 2.1.6 and FP32 tensorflow-cpu 2.21.0 inference; pair comparison exit_code=0,PASS: quant ai-edge-litert 2.1.6 and FP32 tensorflow-cpu 2.21.0 inference; pair comparison exit_code=0,PASS: quant ai-edge-litert 2.1.6 and FP32 tensorflow-cpu 2.21.0 inference; pair comparison exit_code=0,TRUE,https://storage.googleapis.com/download.tensorflow.org/models/deeplabv3_mnv2_dm05_pascal_trainaug_2018_10_01.tar.gz,research/downloads/segmentation/fp32_extracted/deeplabv3_mnv2_dm05_pascal_trainaug/frozen_inference_graph.pb,TensorFlow GraphDef/SavedModel,google-coral/test_data@104342d2d3480b3e66203073dac24f4e2dbb4c41; FP32 checkpoint 2018-10-01,61d836df7b46044ca57ddee4016848d1f54b3a343e64ec4cc0cf057c9b13c993,Apache-2.0 for TensorFlow/Coral model artifacts; PASCAL VOC images retain Flickr/source terms,VERIFIED,"PASCAL VOC 2012 train_aug with ImageNet/MS-COCO pretraining; 21 IDs including background, ignore label 255",UNKNOWN,mean Intersection-over-Union,"quant [1,513,513,3]; FP32 [1,H,W,3] invoked at 513x513","[1,513,513]",RGB uint8; FP32 GraphDef has dynamic spatial input and was invoked at 513x513; public quantized artifact fixes 513x513 with scale=1/128 and zero_point=128,Task-specific postprocessing is defined by the corresponding pipeline config,21 PASCAL VOC category IDs including background; ignore label 255,FALSE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/segmentation/deeplabv3_mnv2_dm05_pascal_quant.tflite,research/downloads/segmentation/fp32_extracted/deeplabv3_mnv2_dm05_pascal_trainaug/frozen_inference_graph.pb,https://coral.ai/models/semantic-segmentation/,research/evidence/segmentation/sources/coral_semantic_segmentation.html;research/evidence/segmentation/sources/coral_test_data_LICENSE;research/evidence/segmentation/logs/download_checksums.tsv;research/evidence/segmentation/runtime/quantization_summary.json;research/evidence/segmentation/sources/tensorflow_deeplab_model_zoo.md;research/evidence/segmentation/sources/pascal_voc2012_dataset.html;research/evidence/segmentation/sources/pascal_voc_segmentation_labels.txt;research/evidence/segmentation/runtime/SG07-public-quantized-01.json;research/evidence/segmentation/runtime/SG07-fp32-01.json;research/evidence/segmentation/runtime/pair_compatibility.json,curl -fL 'https://coral.ai/models/semantic-segmentation/',200,ELIGIBLE,P0,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
SG08,SG08-public-quantized-01,EdgeTPU-DeepLab-slim width 0.75 Cityscapes,semantic_segmentation,EdgeTPU-MobileNet width-0.75 depthwise CNN + slim DeepLab head,TensorFlow 1 frozen GraphDef / TFLite,https://github.com/tensorflow/models/tree/archive/research/deeplab,Apache-2.0 for TensorFlow/Coral model artifacts; Cityscapes dataset is limited to agreed non-commercial terms,Apache-2.0 for TensorFlow/Coral model artifacts; Cityscapes dataset is limited to agreed non-commercial terms,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://coral.ai/models/semantic-segmentation/,research/downloads/segmentation/deeplab_mobilenet_edgetpu_slim_cityscapes_quant.tflite,TFLite FlatBuffer,google-coral/test_data@104342d2d3480b3e66203073dac24f4e2dbb4c41; FP32 checkpoint 2020-03-09,322fe29945cb1693f426952f2a330ab67ea5678545ddee6f2548e0cfd0ca03e5,Apache-2.0 for TensorFlow/Coral model artifacts; Cityscapes dataset is limited to agreed non-commercial terms,FULL_INTEGER,per-tensor verified (208 quantized tensors; no multi-scale tensor),"weights, activations, and input; integer ArgMax output",uint8 (plus int32 bias),uint8,uint8; scale=0.0078125; zero_point=128,int64 category IDs (paired FP32 GraphDef outputs int32 category IDs),TRUE,PASS: quant ai-edge-litert 2.1.6 and FP32 tensorflow-cpu 2.21.0 inference; pair comparison exit_code=0,PASS: quant ai-edge-litert 2.1.6 and FP32 tensorflow-cpu 2.21.0 inference; pair comparison exit_code=0,PASS: quant ai-edge-litert 2.1.6 and FP32 tensorflow-cpu 2.21.0 inference; pair comparison exit_code=0,TRUE,https://storage.googleapis.com/download.tensorflow.org/models/edgetpu-deeplab-slim_2020_03_09.tar.gz,research/downloads/segmentation/fp32_extracted/edgetpu-deeplab-slim/frozen_inference_graph.pb,TensorFlow GraphDef/SavedModel,google-coral/test_data@104342d2d3480b3e66203073dac24f4e2dbb4c41; FP32 checkpoint 2020-03-09,7ff42b476610d394486fa7cc015c042b74adec234744dd72ff80d3637f94089d,Apache-2.0 for TensorFlow/Coral model artifacts; Cityscapes dataset is limited to agreed non-commercial terms,VERIFIED,Cityscapes train_fine; supplied map has background plus 19 evaluation classes (Coral table describes 28 objects),UNKNOWN,mean Intersection-over-Union,"quant [1,513,513,3]; FP32 [1,H,W,3] invoked at 513x513","[1,513,513]",RGB uint8; FP32 GraphDef has dynamic spatial input and was invoked at 513x513; public quantized artifact fixes 513x513 with scale=1/128 and zero_point=128,Task-specific postprocessing is defined by the corresponding pipeline config,20 supplied output IDs: background plus 19 Cityscapes evaluation classes,FALSE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/segmentation/deeplab_mobilenet_edgetpu_slim_cityscapes_quant.tflite,research/downloads/segmentation/fp32_extracted/edgetpu-deeplab-slim/frozen_inference_graph.pb,https://coral.ai/models/semantic-segmentation/,research/evidence/segmentation/sources/coral_semantic_segmentation.html;research/evidence/segmentation/sources/coral_test_data_LICENSE;research/evidence/segmentation/logs/download_checksums.tsv;research/evidence/segmentation/runtime/quantization_summary.json;research/evidence/segmentation/logs/sg08_fp32_download_checksums.tsv;research/evidence/segmentation/sources/tensorflow_deeplab_model_zoo.md;research/evidence/segmentation/sources/cityscapes_license.html;research/evidence/segmentation/sources/cityscapes_segmentation_labels.txt;research/evidence/segmentation/runtime/SG08-public-quantized-01.json;research/evidence/segmentation/runtime/SG08-fp32-01.json;research/evidence/segmentation/runtime/pair_compatibility.json,curl -fL 'https://coral.ai/models/semantic-segmentation/',200,ELIGIBLE,P1,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
SP01,SP01-public-quantized-01,MLPerf Tiny DS-CNN KWS reference,keyword spotting,depthwise-separable CNN,TensorFlow/Keras SavedModel and TensorFlow Lite,https://github.com/mlcommons/tiny/tree/v1.1/benchmark/training/keyword_spotting,Apache-2.0; repository LICENSE.md SHA-256 0d542e0c8804e39aa7f37eb00da5a762149dc682d7829451287e11b938e94594,Apache-2.0; repository LICENSE.md SHA-256 0d542e0c8804e39aa7f37eb00da5a762149dc682d7829451287e11b938e94594,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://raw.githubusercontent.com/mlcommons/tiny/v1.1/benchmark/training/keyword_spotting/trained_models/kws_ref_model.tflite,research/downloads/speech_kws/mlperf_v1.1_kws_ref_model_int8.tflite,TFLite FlatBuffer,MLCommons Tiny tag v1.1; commit 360bf095d4620057cbda8c73a507b2a61341c669,aeea436800704fce17b17292e4412630ad856e9d777c044c64ef748a880bd0ae,Apache-2.0; repository LICENSE.md SHA-256 0d542e0c8804e39aa7f37eb00da5a762149dc682d7829451287e11b938e94594,full-integer affine TFLite; upstream PTQ artifact used unchanged,per-channel Conv/Depthwise/Dense weights and per-tensor activations (18 per-channel tensors measured),weights; biases; activations; int8 input; int8 output,int8 (bias int32),int8,int8; scale=0.5847029089927673; zero_point=83,int8; scale=0.00390625; zero_point=-128,TRUE,PASS: ai-edge-litert 2.1.4 allocate+invoke for int8 and FP32 TFLite; TensorFlow 2.15.1 SavedModel load+invoke,ai-edge-litert 2.1.4; tensorflow-cpu 2.15.1; Python 3.11.15,ai-edge-litert 2.1.4; tensorflow-cpu 2.15.1; Python 3.11.15,TRUE,https://raw.githubusercontent.com/mlcommons/tiny/v1.1/benchmark/training/keyword_spotting/trained_models/kws_ref_model_float32.tflite,research/downloads/speech_kws/mlperf_v1.1_kws_ref_model_float32.tflite,TFLite FlatBuffer; upstream float32-reference runtime with float32 I/O/activations and 5 per-channel INT8 constant tensors (HYBRID_OR_WEIGHT_ONLY observed),MLCommons Tiny tag v1.1; commit 360bf095d4620057cbda8c73a507b2a61341c669,e5004c6f1012246e33fa068d8488325538e0444073cd361f5a7edb40c73f12d2,Apache-2.0; repository LICENSE.md SHA-256 0d542e0c8804e39aa7f37eb00da5a762149dc682d7829451287e11b938e94594,VERIFIED,Google Speech Commands v2,UNKNOWN,task-specific quality metric,"[1,49,10,1] (batch signature is dynamic; MCU batch fixed to 1)","[1,12]",16 kHz mono; 1 s; 30 ms window; 20 ms stride; 10 MFCC; 49 frames,Task-specific postprocessing is defined by the corresponding pipeline config,ordered 12 classes: Down; Go; Left; No; Off; On; Right; Stop; Up; Yes; Silence; Unknown,TRUE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/speech_kws/mlperf_v1.1_kws_ref_model_int8.tflite,research/downloads/speech_kws/mlperf_v1.1_kws_ref_model_float32.tflite,https://raw.githubusercontent.com/mlcommons/tiny/v1.1/benchmark/training/keyword_spotting/trained_models/kws_ref_model.tflite,research/evidence/speech_kws/mlperf_ds_cnn_quant_runtime_v2.json; research/evidence/speech_kws/mlperf_ds_cnn_fp32_runtime_v2.json; research/evidence/speech_kws/logs/mlperf_ds_cnn_fp32_savedmodel_runtime.log; research/evidence/speech_kws/logs/mlperf_kws_evidence.log,curl -fL 'https://raw.githubusercontent.com/mlcommons/tiny/v1.1/benchmark/training/keyword_spotting/trained_models/kws_ref_model.tflite',200,ELIGIBLE,P0,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
SP02,SP02-public-quantized-01,MLPerf Tiny streaming wakeword 1D DS-CNN,keyword spotting,temporal depthwise-separable CNN,TensorFlow/Keras H5 and TensorFlow Lite,https://github.com/mlcommons/tiny/tree/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/streaming_wakeword,Apache-2.0; MLCommons Tiny repository,Apache-2.0; MLCommons Tiny repository,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/streaming_wakeword/trained_models/str_ww_ref_model.tflite,research/downloads/speech_kws/mlperf_4addd0f_str_ww_ref_model_int8.tflite,TFLite FlatBuffer,commit 4addd0fa08d216e20637637874e084895f289da4 (2026-07-06 commit date),3af8550895ba7d5c584277102b5075c52dcfa63ba9d2b2240f37c4e6abd5dd2b,Apache-2.0; MLCommons Tiny repository,full-integer affine TFLite public artifact; upstream training/conversion may use QAT and representative calibration,per-channel temporal depthwise/pointwise/Dense weights and per-tensor activations (16 per-channel tensors measured),weights; biases; activations; int8 input; int8 output,int8 (bias int32),int8,int8; scale=0.003701042616739869; zero_point=-128,int8; scale=0.00390625; zero_point=-128,TRUE,PASS: ai-edge-litert 2.1.4 int8 allocate+invoke; TensorFlow 2.15.1 FP32 H5 load+invoke; finite outputs and matching zero-input top-1,ai-edge-litert 2.1.4; tensorflow-cpu 2.15.1; Python 3.11.15,ai-edge-litert 2.1.4; tensorflow-cpu 2.15.1; Python 3.11.15,TRUE,https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/streaming_wakeword/trained_models/str_ww_ref_model.h5,research/downloads/speech_kws/mlperf_4addd0f_str_ww_ref_model_fp32.h5,Keras H5,commit 4addd0fa08d216e20637637874e084895f289da4 (2026-07-06 commit date),b0f267a8ba0bcb911c1098229c32fac21996e4191c9c60d1fda80adaa70a8add,Apache-2.0; MLCommons Tiny repository,VERIFIED,Google Speech Commands v0.02 plus MUSAN for streaming/noise evaluation,UNKNOWN,task-specific quality metric,"[1,30,1,40] (batch signature is dynamic; MCU batch fixed to 1)","[1,3]","16 kHz mono; 1 s; preemphasis 1-2^-5; 64 ms nonperiodic Hamming; 32 ms stride; power spectrum; 40-bin 0-8 kHz LFBE; log10 and clip/scale to [0,1]; 30 frames",Task-specific postprocessing is defined by the corresponding pipeline config,ordered 3 classes: Marvin; Silence; Unknown,TRUE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/speech_kws/mlperf_4addd0f_str_ww_ref_model_int8.tflite,research/downloads/speech_kws/mlperf_4addd0f_str_ww_ref_model_fp32.h5,https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/streaming_wakeword/trained_models/str_ww_ref_model.tflite,research/evidence/speech_kws/mlperf_1d_dscnn_quant_runtime.json; research/evidence/speech_kws/logs/mlperf_1d_dscnn_fp32_runtime.log; research/evidence/speech_kws/logs/mlperf_1d_dscnn_preprocessing.log,curl -fL 'https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/streaming_wakeword/trained_models/str_ww_ref_model.tflite',200,ELIGIBLE,P0,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
SP08,SP08-public-quantized-01,TensorFlow Lite Micro micro_speech tiny_conv,keyword spotting,tiny depthwise convolution plus fully connected,TensorFlow frozen GraphDef and TensorFlow Lite,https://android.googlesource.com/platform/external/tensorflow/+/4ed0453e335/tensorflow/lite/micro/examples/micro_speech/train/,Apache-2.0; source LICENSE SHA-256 37dbed59a6aef0803f24a7dac146679491c460faa8bc3834974d786acf56a40f,Apache-2.0; source LICENSE SHA-256 37dbed59a6aef0803f24a7dac146679491c460faa8bc3834974d786acf56a40f,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://storage.googleapis.com/download.tensorflow.org/models/tflite/micro/micro_speech_2020_04_13.zip#micro_speech/models/model.tflite,research/downloads/speech_kws/micro_speech_2020_04_13/micro_speech/models/model.tflite,TFLite FlatBuffer inside ZIP,micro_speech_2020_04_13 release archive; TensorFlow source snapshot 4ed0453e335,454779dcfea05290759256178162fa86eb17642e7c9cac0de8ea78e1693cee00,Apache-2.0; source LICENSE SHA-256 37dbed59a6aef0803f24a7dac146679491c460faa8bc3834974d786acf56a40f,full-integer affine TFLite; README says strictly int8 but measured legacy artifact tensors are uint8/int32,per-tensor weights and activations (0 per-channel tensors measured),weights; biases; activations; uint8 feature input; uint8 probability output,uint8 (bias int32),uint8,uint8; scale=0.10196070373058319; zero_point=0,uint8; scale=0.00390625; zero_point=0,TRUE,PASS: ai-edge-litert 2.1.4 quant allocate+invoke; TensorFlow 2.15.1 FP32 GraphDef load+invoke; internal-feature pair check PASS,ai-edge-litert 2.1.4; tensorflow-cpu 2.15.1; Python 3.11.15,ai-edge-litert 2.1.4; tensorflow-cpu 2.15.1; Python 3.11.15,TRUE,https://storage.googleapis.com/download.tensorflow.org/models/tflite/micro/micro_speech_2020_04_13.zip#micro_speech/models/model.pb,research/downloads/speech_kws/micro_speech_2020_04_13/micro_speech/models/model.pb,TensorFlow GraphDef/SavedModel,micro_speech_2020_04_13 release archive; TensorFlow source snapshot 4ed0453e335,61f3f6db6a0eb11293c8b64d9916da9095ad30d9042db2227b86824d29dfec4e,Apache-2.0; source LICENSE SHA-256 37dbed59a6aef0803f24a7dac146679491c460faa8bc3834974d786acf56a40f,VERIFIED,Google Speech Commands v2,UNKNOWN,task-specific quality metric,"quant core [1,49,40,1]; FP32 graph accepts scalar WAV bytes and exposes matching internal Reshape_2 [1,49,40,1]","[1,4]",16 kHz mono; 1 s; microfrontend spectrogram; 30 ms feature slice; 20 ms stride; 49 slices; 40 bins,Task-specific postprocessing is defined by the corresponding pipeline config,ordered 4 classes: silence; unknown; yes; no,FALSE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/speech_kws/micro_speech_2020_04_13/micro_speech/models/model.tflite,research/downloads/speech_kws/micro_speech_2020_04_13/micro_speech/models/model.pb,https://storage.googleapis.com/download.tensorflow.org/models/tflite/micro/micro_speech_2020_04_13.zip#micro_speech/models/model.tflite,research/evidence/speech_kws/tensorflow_micro_speech_quant_runtime.json; research/evidence/speech_kws/tensorflow_micro_speech_pair_runtime.json; research/evidence/speech_kws/sources/micro_model_settings.h; research/evidence/speech_kws/sources/micro_model_settings.cc,curl -fL 'https://storage.googleapis.com/download.tensorflow.org/models/tflite/micro/micro_speech_2020_04_13.zip#micro_speech/models/model.tflite',200,ELIGIBLE,P0,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
SP09,SP09-public-quantized-01,Arm ML-Zoo clustered DS-CNN Large INT8,keyword spotting,clustered depthwise-separable CNN,TensorFlow Lite; upstream Keras and TensorFlow Model Optimization Toolkit,https://github.com/Arm-Examples/ML-zoo/tree/68b5fbc77ed28e67b2efc915997ea4477c1d9d5b/models/keyword_spotting/ds_cnn_large,Apache-2.0 in artifact definition and repository LICENSE,Apache-2.0 in artifact definition and repository LICENSE,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://github.com/ARM-software/ML-zoo/raw/68b5fbc77ed28e67b2efc915997ea4477c1d9d5b/models/keyword_spotting/ds_cnn_large/tflite_clustered_int8/ds_cnn_clustered_int8.tflite,research/downloads/speech_kws/arm_ml_zoo_ds_cnn_clustered_int8.tflite,TFLite FlatBuffer,Arm ML-Zoo commit 68b5fbc77ed28e67b2efc915997ea4477c1d9d5b,1c254bb3eee70df761cd6bc1320ccad31979ec10b6841b385fdea8304daf6dfe,Apache-2.0 in artifact definition and repository LICENSE,clustered (32 clusters; kmeans++) retrained FP32 then upstream post-training full-integer quantization,per-channel Conv/Depthwise/Dense weights and per-tensor activations (22 per-channel tensors measured),weights; biases; activations; int8 input; int8 output,int8 (bias int32),int8,int8; scale=1.1071635484695435; zero_point=95,int8; scale=0.00390625; zero_point=-128,TRUE,PASS: ai-edge-litert 2.1.4 allocate+invoke on quant and FP32; finite outputs; same zero-input top-1; identical measured operator sequence,ai-edge-litert 2.1.4; Python 3.11.15,ai-edge-litert 2.1.4; Python 3.11.15,TRUE,https://github.com/ARM-software/ML-zoo/raw/68b5fbc77ed28e67b2efc915997ea4477c1d9d5b/models/keyword_spotting/ds_cnn_large/tflite_clustered_fp32/ds_cnn_clustered_fp32.tflite,research/downloads/speech_kws/arm_ml_zoo_ds_cnn_clustered_fp32.tflite,TFLite FlatBuffer,Arm ML-Zoo commit 68b5fbc77ed28e67b2efc915997ea4477c1d9d5b,8bee3734e58c71177ab44a9579283bae5c9844296e122507b85fba36c22ca658,Apache-2.0 in artifact definition and repository LICENSE,VERIFIED,Google Speech Commands,UNKNOWN,task-specific quality metric,"[1,1,49,10]","[1,12]","16 kHz mono; 1 s; processed MFCC [49,10]; original Arm recipe uses 40 ms window and 20 ms stride",Task-specific postprocessing is defined by the corresponding pipeline config,ordered 12 classes from Arm provenance: silence; unknown; yes; no; up; down; left; right; on; off; stop; go,FALSE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/speech_kws/arm_ml_zoo_ds_cnn_clustered_int8.tflite,research/downloads/speech_kws/arm_ml_zoo_ds_cnn_clustered_fp32.tflite,https://github.com/ARM-software/ML-zoo/raw/68b5fbc77ed28e67b2efc915997ea4477c1d9d5b/models/keyword_spotting/ds_cnn_large/tflite_clustered_int8/ds_cnn_clustered_int8.tflite,research/evidence/speech_kws/arm_ml_zoo_ds_cnn_clustered_int8_runtime.json; research/evidence/speech_kws/arm_ml_zoo_ds_cnn_clustered_fp32_runtime.json; research/evidence/speech_kws/sources/arm_ds_cnn_clustered_int8_definition.yaml; research/evidence/speech_kws/sources/arm_ds_cnn_clustered_fp32_definition.yaml,curl -fL 'https://github.com/ARM-software/ML-zoo/raw/68b5fbc77ed28e67b2efc915997ea4477c1d9d5b/models/keyword_spotting/ds_cnn_large/tflite_clustered_int8/ds_cnn_clustered_int8.tflite',200,ELIGIBLE,P0,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
VC01,VC01-public-quantized-01,MLPerf Tiny Visual Wake Words MobileNetV1 0.25,vision_classification,Depthwise CNN / MobileNetV1,TensorFlow/Keras -> TFLite,https://github.com/mlcommons/tiny/tree/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/visual_wake_words,Apache-2.0,Apache-2.0,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/visual_wake_words/trained_models/vww_96_int8.tflite,research/downloads/vision/VC01/vww_96_int8.tflite,TFLite FlatBuffer,4addd0fa08d216e20637637874e084895f289da4,597a384c8c2c8a1276f04702f25013b7838f2f814f1ca7c174d295b73e3d6b7b,Apache-2.0 (repository license),upstream static full-integer TFLite,per-channel weights / per-tensor activations (scale-vector inspection),full graph including integer input/output,int8,int8,int8 scale=0.003921568859 zp=-128,int8 scale=0.00390625 zp=-128,TRUE,PASS: tflite-runtime 2.14.0 allocate+zero-input invoke; research/evidence/vision/runtime_results/VC01.json; FP32 research/evidence/vision/runtime_results/VC01_fp32.json,PASS: tflite-runtime 2.14.0 allocate+zero-input invoke; research/evidence/vision/runtime_results/VC01.json; FP32 research/evidence/vision/runtime_results/VC01_fp32.json,PASS: tflite-runtime 2.14.0 allocate+zero-input invoke; research/evidence/vision/runtime_results/VC01.json; FP32 research/evidence/vision/runtime_results/VC01_fp32.json,TRUE,https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/visual_wake_words/trained_models/vww_96_float.tflite,research/downloads/vision/VC01/vww_96_float.tflite,TFLite FlatBuffer,4addd0fa08d216e20637637874e084895f289da4,115bbc094d2119561320a21f01b6500a18bea8cc8589282ab007097bec8af38c,Apache-2.0,VERIFIED,Visual Wake Words derived from MS COCO (person / no-person),COCO image-specific terms; VWW-derived labels; dataset not redistributed,Top-1 accuracy,"[N,96,96,3] NHWC (allocated N=1)","[1,2]",RGB resize 96x96; source training divides pixels by 255; int8 input scale maps byte domain,Task-specific postprocessing is defined by the corresponding pipeline config,Two Visual Wake Words classes: person and non-person,FALSE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/vision/VC01/vww_96_int8.tflite,research/downloads/vision/VC01/vww_96_float.tflite,https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/visual_wake_words/trained_models/vww_96_int8.tflite,research/evidence/vision/sources/mlcommons/convert_vww.py | research/evidence/vision/sources/mlcommons/vww_model.py | research/evidence/vision/runtime_results/VC01.json,curl -fL 'https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/visual_wake_words/trained_models/vww_96_int8.tflite',200,ELIGIBLE,P0,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
VC02,VC02-public-quantized-01,MLPerf Tiny CIFAR-10 ResNet8,vision_classification,Residual CNN / ResNet8,TensorFlow/Keras -> TFLite,https://github.com/mlcommons/tiny/tree/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/image_classification,Apache-2.0,Apache-2.0,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/image_classification/trained_models/pretrainedResnet_quant.tflite,research/downloads/vision/VC02/pretrainedResnet_quant.tflite,TFLite FlatBuffer,4addd0fa08d216e20637637874e084895f289da4,3c002613d1b2475eb51dd78dfb85a546c8ae658dee71cf6ade43b022fe205415,Apache-2.0 (repository license),upstream static full-integer TFLite,per-channel weights / per-tensor activations (scale-vector inspection),full graph including integer input/output,int8,int8,int8 scale=1.0 zp=-128,int8 scale=0.00390625 zp=-128,TRUE,PASS: tflite-runtime 2.14.0 allocate+zero-input invoke; research/evidence/vision/runtime_results/VC02.json; FP32 research/evidence/vision/runtime_results/VC02_fp32.json,PASS: tflite-runtime 2.14.0 allocate+zero-input invoke; research/evidence/vision/runtime_results/VC02.json; FP32 research/evidence/vision/runtime_results/VC02_fp32.json,PASS: tflite-runtime 2.14.0 allocate+zero-input invoke; research/evidence/vision/runtime_results/VC02.json; FP32 research/evidence/vision/runtime_results/VC02_fp32.json,TRUE,https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/image_classification/trained_models/pretrainedResnet.tflite,research/downloads/vision/VC02/pretrainedResnet.tflite,TFLite FlatBuffer,4addd0fa08d216e20637637874e084895f289da4,b5c0046d6e0328b4956afd6baa29555a29b1f1c65bdd45aaed75b7cd484d9f79,Apache-2.0,VERIFIED,CIFAR-10,UNKNOWN_NOT_STATED_IN_PINNED_MODEL_SOURCE; dataset bytes not redistributed,Top-1 accuracy,"[N,32,32,3] NHWC (allocated N=1)","[1,10]",32x32 RGB; converter/model sources preserve CIFAR-10 input and do not embed a normalization layer,Task-specific postprocessing is defined by the corresponding pipeline config,Ten CIFAR-10 classes in the dataset's canonical index order,FALSE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/vision/VC02/pretrainedResnet_quant.tflite,research/downloads/vision/VC02/pretrainedResnet.tflite,https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/image_classification/trained_models/pretrainedResnet_quant.tflite,research/evidence/vision/sources/mlcommons/model_converter.py | research/evidence/vision/sources/mlcommons/keras_model.py | research/evidence/vision/runtime_results/VC02.json,curl -fL 'https://raw.githubusercontent.com/mlcommons/tiny/4addd0fa08d216e20637637874e084895f289da4/benchmark/training/image_classification/trained_models/pretrainedResnet_quant.tflite',200,ELIGIBLE,P0,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
VC03,VC03-public-quantized-01,TensorFlow MobileNetV1 alpha=0.25 224,vision_classification,Depthwise CNN / MobileNetV1,TensorFlow Slim -> TFLite,https://github.com/tensorflow/models/blob/4d7bdd8c170ee90850f2f9ccef0f6d19b817de35/research/slim/nets/mobilenet_v1.md,Apache-2.0,Apache-2.0,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.25_224_quant.tgz,research/downloads/vision/VC03/mobilenet_v1_0.25_224_quant.tflite,TFLite FlatBuffer inside TGZ,2018-08-02,ce8f61fe3e4e29560c4ea8038e4b91eb354be330fc428348b9645bd1e7de03bf,Apache-2.0 (TensorFlow models distribution),upstream quantization-aware training/FakeQuant then fully-quantized TFLite,per-tensor weights and activations (all inspected scale vectors length 1),full graph including uint8 input/output,uint8,uint8/int32 accumulator,uint8 scale=0.0078125 zp=128,uint8 scale=0.00390625 zp=0,TRUE,PASS: tflite-runtime 2.14.0 allocate+zero-input invoke; research/evidence/vision/runtime_results/VC03.json; FP32 research/evidence/vision/runtime_results/VC03_fp32.json,PASS: tflite-runtime 2.14.0 allocate+zero-input invoke; research/evidence/vision/runtime_results/VC03.json; FP32 research/evidence/vision/runtime_results/VC03_fp32.json,PASS: tflite-runtime 2.14.0 allocate+zero-input invoke; research/evidence/vision/runtime_results/VC03.json; FP32 research/evidence/vision/runtime_results/VC03_fp32.json,TRUE,https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.25_224.tgz,research/downloads/vision/VC03/mobilenet_v1_0.25_224.tflite,TFLite FlatBuffer,2018-08-02,809472d94a9e80e6cce276a5fe14e47a2868c2afac1fd2dc25a20fbf45a5efa3,Apache-2.0,VERIFIED,ImageNet 2012 validation; 1000 classes plus background output slot,ImageNet terms; dataset not redistributed,Top-1 accuracy,"[1,224,224,3] NHWC","[1,1001]",RGB 224x224; quant input dequantization (u8-128)/128; ImageNet label order with background index,Task-specific postprocessing is defined by the corresponding pipeline config,"ImageNet 2012 1000 classes plus background output slot, using the paired artifact's official index order",FALSE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/vision/VC03/mobilenet_v1_0.25_224_quant.tflite,research/downloads/vision/VC03/mobilenet_v1_0.25_224.tflite,https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.25_224_quant.tgz,research/evidence/vision/sources/tensorflow_models/mobilenet_v1.md | research/evidence/vision/runtime_results/VC03.json | research/evidence/vision/logs/downloads/VC03/extracted_quantized.sha256.txt,curl -fL 'https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.25_224_quant.tgz',200,ELIGIBLE,P0,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
VC04,VC04-public-quantized-01,TensorFlow MobileNetV1 alpha=0.5 224,vision_classification,Depthwise CNN / MobileNetV1,TensorFlow Slim -> TFLite,https://github.com/tensorflow/models/blob/4d7bdd8c170ee90850f2f9ccef0f6d19b817de35/research/slim/nets/mobilenet_v1.md,Apache-2.0,Apache-2.0,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.5_224_quant.tgz,research/downloads/vision/VC04/mobilenet_v1_0.5_224_quant.tflite,TFLite FlatBuffer inside TGZ,2018-08-02,b14dcdda08a2a8746977ef312504b86e4205c10508de5b50816fd380f6a679db,Apache-2.0 (TensorFlow models distribution),upstream quantization-aware training/FakeQuant then fully-quantized TFLite,per-tensor weights and activations (all inspected scale vectors length 1),full graph including uint8 input/output,uint8,uint8/int32 accumulator,uint8 scale=0.0078125 zp=128,uint8 scale=0.00390625 zp=0,TRUE,PASS: tflite-runtime 2.14.0 allocate+zero-input invoke; research/evidence/vision/runtime_results/VC04.json; FP32 research/evidence/vision/runtime_results/VC04_fp32.json,PASS: tflite-runtime 2.14.0 allocate+zero-input invoke; research/evidence/vision/runtime_results/VC04.json; FP32 research/evidence/vision/runtime_results/VC04_fp32.json,PASS: tflite-runtime 2.14.0 allocate+zero-input invoke; research/evidence/vision/runtime_results/VC04.json; FP32 research/evidence/vision/runtime_results/VC04_fp32.json,TRUE,https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.5_224.tgz,research/downloads/vision/VC04/mobilenet_v1_0.5_224.tflite,TFLite FlatBuffer,2018-08-02,fb12f819f74cdc9910b69e9c282b953b893d8578ed3d64320c0786c28a8f0d50,Apache-2.0,VERIFIED,ImageNet 2012 validation; 1000 classes plus background output slot,ImageNet terms; dataset not redistributed,Top-1 accuracy,"[1,224,224,3] NHWC","[1,1001]",RGB 224x224; quant input dequantization (u8-128)/128; ImageNet label order with background index,Task-specific postprocessing is defined by the corresponding pipeline config,"ImageNet 2012 1000 classes plus background output slot, using the paired artifact's official index order",FALSE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/vision/VC04/mobilenet_v1_0.5_224_quant.tflite,research/downloads/vision/VC04/mobilenet_v1_0.5_224.tflite,https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.5_224_quant.tgz,research/evidence/vision/sources/tensorflow_models/mobilenet_v1.md | research/evidence/vision/runtime_results/VC04.json | research/evidence/vision/logs/downloads/VC04/extracted_quantized.sha256.txt,curl -fL 'https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.5_224_quant.tgz',200,ELIGIBLE,P0,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
VC05,VC05-public-quantized-01,ST MobileNetV2 alpha=0.35 224,vision_classification,Depthwise CNN + residual / MobileNetV2,TensorFlow/Keras -> TFLite,https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/1423c78953a830903485135febe1dd98ff31aed8/image_classification/mobilenetv2,Apache-2.0,Apache-2.0,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://media.githubusercontent.com/media/STMicroelectronics/stm32ai-modelzoo/1423c78953a830903485135febe1dd98ff31aed8/image_classification/mobilenetv2/Public_pretrainedmodel_public_dataset/ImageNet/mobilenetv2_a035_224/mobilenetv2_a035_224_int8.tflite,research/downloads/vision/VC05/mobilenetv2_a035_224_int8.tflite,TFLite FlatBuffer,1423c78953a830903485135febe1dd98ff31aed8,194192ddbf615a2c2ffb842a4370e8ff88fa6547958137e97cb09aaf44747067,Apache-2.0 (artifact-directory LICENSE.md),upstream static INT8 TFLite,per-channel weights / per-tensor activations (scale-vector inspection),integer body and uint8 input; float32 output boundary,int8,int8 with int32 accumulators,uint8 scale=0.007843137719 zp=127,float32,FALSE,PASS: tflite-runtime 2.14.0 quant invoke + Keras 3.8.0 torch-backend FP32 load/invoke; research/evidence/vision/runtime_results/VC05.json; research/evidence/vision/runtime_results/VC05_fp32.json,PASS: tflite-runtime 2.14.0 quant invoke + Keras 3.8.0 torch-backend FP32 load/invoke; research/evidence/vision/runtime_results/VC05.json; research/evidence/vision/runtime_results/VC05_fp32.json,PASS: tflite-runtime 2.14.0 quant invoke + Keras 3.8.0 torch-backend FP32 load/invoke; research/evidence/vision/runtime_results/VC05.json; research/evidence/vision/runtime_results/VC05_fp32.json,TRUE,https://raw.githubusercontent.com/STMicroelectronics/stm32ai-modelzoo/1423c78953a830903485135febe1dd98ff31aed8/image_classification/mobilenetv2/Public_pretrainedmodel_public_dataset/ImageNet/mobilenetv2_a035_224/mobilenetv2_a035_224.keras,research/downloads/vision/VC05/mobilenetv2_a035_224.keras,Keras v3,1423c78953a830903485135febe1dd98ff31aed8,a8e532a5bb25e8d29435ef159f6fbb81cc1358ec5716c98d5120794c2334d0cd,Apache-2.0,VERIFIED,ImageNet,ImageNet terms; dataset not redistributed,Top-1 accuracy,"[N,224,224,3] NHWC (allocated N=1)","[1,1000]","RGB resize 224x224; source config rescaling scale=1/127.5, offset=-1",Task-specific postprocessing is defined by the corresponding pipeline config,ImageNet 1000 classes using the paired ST artifact's class order,FALSE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/vision/VC05/mobilenetv2_a035_224_int8.tflite,research/downloads/vision/VC05/mobilenetv2_a035_224.keras,https://media.githubusercontent.com/media/STMicroelectronics/stm32ai-modelzoo/1423c78953a830903485135febe1dd98ff31aed8/image_classification/mobilenetv2/Public_pretrainedmodel_public_dataset/ImageNet/mobilenetv2_a035_224/mobilenetv2_a035_224_int8.tflite,research/evidence/vision/sources/stm32ai/mobilenetv2_README.md | research/evidence/vision/sources/stm32ai/mobilenetv2_a035_224_config.yaml | research/evidence/vision/runtime_results/VC05.json | research/evidence/vision/runtime_results/VC05_fp32.json,curl -fL 'https://media.githubusercontent.com/media/STMicroelectronics/stm32ai-modelzoo/1423c78953a830903485135febe1dd98ff31aed8/image_classification/mobilenetv2/Public_pretrainedmodel_public_dataset/ImageNet/mobilenetv2_a035_224/mobilenetv2_a035_224_int8.tflite',200,ELIGIBLE,P0,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
VC06,VC06-public-quantized-01,ST MobileNetV2 alpha=0.50 PyTorch 224 QDQ,vision_classification,Depthwise CNN + residual / MobileNetV2,PyTorch -> ONNX,https://github.com/STMicroelectronics/stm32ai-modelzoo/tree/1423c78953a830903485135febe1dd98ff31aed8/image_classification/mobilenetv2_pt,Apache-2.0,Apache-2.0,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://media.githubusercontent.com/media/STMicroelectronics/stm32ai-modelzoo/1423c78953a830903485135febe1dd98ff31aed8/image_classification/mobilenetv2_pt/Public_pretrainedmodel_public_dataset/Imagenet/mobilenetv2_a050_pt_224/mobilenetv2_a050_pt_224_qdq_int8.onnx,research/downloads/vision/VC06/mobilenetv2_a050_pt_224_qdq_int8.onnx,ONNX QDQ,1423c78953a830903485135febe1dd98ff31aed8,4f4650e17bb9a65ec3c3f0f042200691e25a852b47a0ed40679cc3aeae11d1d9,Apache-2.0 (artifact-directory LICENSE.md),upstream static QDQ INT8 ONNX,per-channel weights / per-tensor activations (scale initializer inspection),quantized internal Conv path; float32 ONNX input/output boundaries,int8,int8 represented by Q/DQ; float boundaries,float32,float32,FALSE,PASS: ONNX 1.18 checker(full_check)+ORT 1.22.1 zero-input invoke; research/evidence/vision/runtime_results/VC06.json; FP32 research/evidence/vision/runtime_results/VC06_fp32.json,PASS: ONNX 1.18 checker(full_check)+ORT 1.22.1 zero-input invoke; research/evidence/vision/runtime_results/VC06.json; FP32 research/evidence/vision/runtime_results/VC06_fp32.json,PASS: ONNX 1.18 checker(full_check)+ORT 1.22.1 zero-input invoke; research/evidence/vision/runtime_results/VC06.json; FP32 research/evidence/vision/runtime_results/VC06_fp32.json,TRUE,https://media.githubusercontent.com/media/STMicroelectronics/stm32ai-modelzoo/1423c78953a830903485135febe1dd98ff31aed8/image_classification/mobilenetv2_pt/Public_pretrainedmodel_public_dataset/Imagenet/mobilenetv2_a050_pt_224/mobilenetv2_a050_pt_224.onnx,research/downloads/vision/VC06/mobilenetv2_a050_pt_224.onnx,ONNX,1423c78953a830903485135febe1dd98ff31aed8,599b2a454c96a7dd156cb38412cd133413dfdd1a1a135b949c2eaf643a438d21,Apache-2.0,VERIFIED,ImageNet,ImageNet terms; dataset not redistributed,Top-1 accuracy,"[N,3,224,224] NCHW","[1,1000]",ImageNet RGB NCHW 224x224; exact normalization requires ST deployment config confirmation before quality run,Task-specific postprocessing is defined by the corresponding pipeline config,ImageNet 1000 classes using the paired ST artifact's class order,FALSE,UNKNOWN,FALSE,TRUE,UNKNOWN,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/vision/VC06/mobilenetv2_a050_pt_224_qdq_int8.onnx,research/downloads/vision/VC06/mobilenetv2_a050_pt_224.onnx,https://media.githubusercontent.com/media/STMicroelectronics/stm32ai-modelzoo/1423c78953a830903485135febe1dd98ff31aed8/image_classification/mobilenetv2_pt/Public_pretrainedmodel_public_dataset/Imagenet/mobilenetv2_a050_pt_224/mobilenetv2_a050_pt_224_qdq_int8.onnx,research/evidence/vision/sources/stm32ai/mobilenetv2_pt_README.md | research/evidence/vision/sources/stm32ai/mobilenetv2_pt_Public_Imagenet_LICENSE.md | research/evidence/vision/runtime_results/VC06.json | research/evidence/vision/runtime_results/VC06_fp32.json,curl -fL 'https://media.githubusercontent.com/media/STMicroelectronics/stm32ai-modelzoo/1423c78953a830903485135febe1dd98ff31aed8/image_classification/mobilenetv2_pt/Public_pretrainedmodel_public_dataset/Imagenet/mobilenetv2_a050_pt_224/mobilenetv2_a050_pt_224_qdq_int8.onnx',200,ELIGIBLE,P1,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
VC09,VC09-public-quantized-01,ONNX Model Zoo SqueezeNet 1.0 INT8,vision_classification,Fire-module CNN / SqueezeNet 1.0,ONNX (upstream converted model),https://github.com/onnx/models/tree/4f43949841cb55a0b98dc8fcd045431ccafd9f96/validated/vision/classification/squeezenet,Apache-2.0,Apache-2.0,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://media.githubusercontent.com/media/onnx/models/4f43949841cb55a0b98dc8fcd045431ccafd9f96/validated/vision/classification/squeezenet/model/squeezenet1.0-12-int8.onnx,research/downloads/vision/VC09/squeezenet1.0-12-int8.onnx,ONNX QOperator INT8,4f43949841cb55a0b98dc8fcd045431ccafd9f96; opset 12,3da17dfad1b7ba23c93fac6dbf49f6db78cd42f7519e915a2e27d37c5c0a972b,Apache-2.0 (model-card license),upstream static PTQ via Intel Neural Compressor/ONNX Runtime,per-channel weights / per-tensor activations (scale initializer inspection),QLinearConv/QLinearMatMul internal graph; float32 input/output boundaries,int8,uint8/int8 internal; float32 boundaries,float32,float32,FALSE,PASS: ONNX 1.18 checker(full_check)+ORT 1.22.1 zero-input invoke; research/evidence/vision/runtime_results/VC09.json; FP32 research/evidence/vision/runtime_results/VC09_fp32.json,PASS: ONNX 1.18 checker(full_check)+ORT 1.22.1 zero-input invoke; research/evidence/vision/runtime_results/VC09.json; FP32 research/evidence/vision/runtime_results/VC09_fp32.json,PASS: ONNX 1.18 checker(full_check)+ORT 1.22.1 zero-input invoke; research/evidence/vision/runtime_results/VC09.json; FP32 research/evidence/vision/runtime_results/VC09_fp32.json,TRUE,https://media.githubusercontent.com/media/onnx/models/4f43949841cb55a0b98dc8fcd045431ccafd9f96/validated/vision/classification/squeezenet/model/squeezenet1.0-12.onnx,research/downloads/vision/VC09/squeezenet1.0-12.onnx,ONNX,4f43949841cb55a0b98dc8fcd045431ccafd9f96; opset 12,dec81a8684617770b3cf13fadc1d92565d1d453d23935fc6388b792d99c992bd,Apache-2.0,VERIFIED,ImageNet-1K,ImageNet terms; dataset not redistributed,Top-1 accuracy,"[1,3,224,224] NCHW","[1,1000,1,1]","RGB 224x224; range [0,1], normalize mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225]",Task-specific postprocessing is defined by the corresponding pipeline config,ImageNet-1K 1000 categories in the ONNX Model Zoo synset order,FALSE,UNKNOWN,FALSE,TRUE,UNKNOWN,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/vision/VC09/squeezenet1.0-12-int8.onnx,research/downloads/vision/VC09/squeezenet1.0-12.onnx,https://media.githubusercontent.com/media/onnx/models/4f43949841cb55a0b98dc8fcd045431ccafd9f96/validated/vision/classification/squeezenet/model/squeezenet1.0-12-int8.onnx,research/evidence/vision/sources/onnx_models/squeezenet_README.md | research/evidence/vision/runtime_results/VC09.json | research/evidence/vision/runtime_results/VC09_fp32.json,curl -fL 'https://media.githubusercontent.com/media/onnx/models/4f43949841cb55a0b98dc8fcd045431ccafd9f96/validated/vision/classification/squeezenet/model/squeezenet1.0-12-int8.onnx',200,ELIGIBLE,P1,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
VC11,VC11-public-quantized-01,Google MediaPipe EfficientNet-Lite0 224 INT8,vision_classification,MBConv/Depthwise CNN / EfficientNet-Lite0,TensorFlow Lite / MediaPipe model distribution,https://developers.google.com/edge/mediapipe/solutions/vision/image_classifier,Apache-2.0 artifact metadata,Apache-2.0 artifact metadata,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://storage.googleapis.com/mediapipe-models/image_classifier/efficientnet_lite0/int8/latest/efficientnet_lite0.tflite,research/downloads/vision/VC11/efficientnet_lite0_int8.tflite,TFLite FlatBuffer,latest resolved 2026-08-06; immutable identity is recorded SHA-256,bc2ffe19c1118de0c0c2a9088992da5589722656e0fba81421385300a4a34b16,Apache-2.0 (embedded string in both downloaded TFLite artifacts),static full-integer TFLite distributed by Google,per-channel weights / per-tensor activations (scale-vector inspection),full graph including uint8 input/output,int8,int8/uint8 with int32 accumulators,uint8 (runtime quantization parameters recorded in JSON),uint8,TRUE,PASS: tflite-runtime 2.14.0 quant+FP allocate and zero-input invoke; research/evidence/vision/runtime_results/VC11.json; research/evidence/vision/runtime_results/VC11_fp32.json,PASS: tflite-runtime 2.14.0 quant+FP allocate and zero-input invoke; research/evidence/vision/runtime_results/VC11.json; research/evidence/vision/runtime_results/VC11_fp32.json,PASS: tflite-runtime 2.14.0 quant+FP allocate and zero-input invoke; research/evidence/vision/runtime_results/VC11.json; research/evidence/vision/runtime_results/VC11_fp32.json,TRUE,https://storage.googleapis.com/mediapipe-models/image_classifier/efficientnet_lite0/float32/latest/efficientnet_lite0.tflite,research/downloads/vision/VC11/efficientnet_lite0_float32.tflite,TFLite FlatBuffer,latest resolved 2026-08-06; immutable identity is recorded SHA-256,6c7ab0a6e5dcbf38a8c33b960996a55a3b4300b36a018c4545801de3a3c8bde0,Apache-2.0 artifact metadata,VERIFIED,ImageNet-1K,ImageNet terms; dataset not redistributed,Top-1 accuracy,"[1,224,224,3] NHWC","[1,1000]",RGB resize 224x224; use embedded TFLite metadata/MediaPipe ImageClassifier preprocessing; quant params in runtime JSON,Task-specific postprocessing is defined by the corresponding pipeline config,ImageNet-1K 1000 categories from the embedded MediaPipe TFLite metadata,FALSE,UNKNOWN,FALSE,UNKNOWN,TRUE,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/vision/VC11/efficientnet_lite0_int8.tflite,research/downloads/vision/VC11/efficientnet_lite0_float32.tflite,https://storage.googleapis.com/mediapipe-models/image_classifier/efficientnet_lite0/int8/latest/efficientnet_lite0.tflite,research/evidence/vision/sources/mediapipe/image_classifier.html | research/evidence/vision/logs/VC11_embedded_license.stdout.log | research/evidence/vision/runtime_results/VC11.json | research/evidence/vision/runtime_results/VC11_fp32.json,curl -fL 'https://storage.googleapis.com/mediapipe-models/image_classifier/efficientnet_lite0/int8/latest/efficientnet_lite0.tflite',200,ELIGIBLE,P0,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
VC12,VC12-public-quantized-01,ONNX Model Zoo MobileNetV2 1.0 INT8,vision_classification,Depthwise CNN + residual / MobileNetV2 1.0,ONNX (upstream converted model),https://github.com/onnx/models/tree/4f43949841cb55a0b98dc8fcd045431ccafd9f96/validated/vision/classification/mobilenet,Apache-2.0,Apache-2.0,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://media.githubusercontent.com/media/onnx/models/4f43949841cb55a0b98dc8fcd045431ccafd9f96/validated/vision/classification/mobilenet/model/mobilenetv2-12-int8.onnx,research/downloads/vision/VC12/mobilenetv2-12-int8.onnx,ONNX QOperator INT8,4f43949841cb55a0b98dc8fcd045431ccafd9f96; opset 12,cc028fe6cae7bc11a4ff53cfc9b79c920e8be65ce33a904ec3e2a8f66d77f95f,Apache-2.0 (model-card license),upstream static PTQ via Intel Neural Compressor/ONNX Runtime,per-channel weights / per-tensor activations (scale initializer inspection),QLinearConv/QLinearMatMul internal graph; float32 input/output boundaries,int8,uint8/int8 internal; float32 boundaries,float32,float32,FALSE,PASS: ONNX 1.18 checker(full_check)+ORT 1.22.1 zero-input invoke; research/evidence/vision/runtime_results/VC12.json; FP32 research/evidence/vision/runtime_results/VC12_fp32.json,PASS: ONNX 1.18 checker(full_check)+ORT 1.22.1 zero-input invoke; research/evidence/vision/runtime_results/VC12.json; FP32 research/evidence/vision/runtime_results/VC12_fp32.json,PASS: ONNX 1.18 checker(full_check)+ORT 1.22.1 zero-input invoke; research/evidence/vision/runtime_results/VC12.json; FP32 research/evidence/vision/runtime_results/VC12_fp32.json,TRUE,https://media.githubusercontent.com/media/onnx/models/4f43949841cb55a0b98dc8fcd045431ccafd9f96/validated/vision/classification/mobilenet/model/mobilenetv2-12.onnx,research/downloads/vision/VC12/mobilenetv2-12.onnx,ONNX,4f43949841cb55a0b98dc8fcd045431ccafd9f96; opset 12,c0c3f76d93fa3fd6580652a45618618a220fced18babf65774ed169de0432ad5,Apache-2.0,VERIFIED,ImageNet-1K,ImageNet terms; dataset not redistributed,Top-1 accuracy,"[N,3,224,224] NCHW","[1,1000]","RGB 224x224; range [0,1], normalize mean=[0.485,0.456,0.406], std=[0.229,0.224,0.225]",Task-specific postprocessing is defined by the corresponding pipeline config,ImageNet-1K 1000 categories in the ONNX Model Zoo synset order,FALSE,UNKNOWN,FALSE,TRUE,UNKNOWN,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/vision/VC12/mobilenetv2-12-int8.onnx,research/downloads/vision/VC12/mobilenetv2-12.onnx,https://media.githubusercontent.com/media/onnx/models/4f43949841cb55a0b98dc8fcd045431ccafd9f96/validated/vision/classification/mobilenet/model/mobilenetv2-12-int8.onnx,research/evidence/vision/sources/onnx_models/mobilenet_README.md | research/evidence/vision/runtime_results/VC12.json | research/evidence/vision/runtime_results/VC12_fp32.json,curl -fL 'https://media.githubusercontent.com/media/onnx/models/4f43949841cb55a0b98dc8fcd045431ccafd9f96/validated/vision/classification/mobilenet/model/mobilenetv2-12-int8.onnx',200,ELIGIBLE,P1,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.
VC13,VC13-public-quantized-01,ONNX Model Zoo ShuffleNetV2 x1.0 INT8,vision_classification,Shuffle/channel-split CNN / ShuffleNetV2 x1.0,ONNX (upstream converted model),https://github.com/onnx/models/tree/4f43949841cb55a0b98dc8fcd045431ccafd9f96/validated/vision/classification/shufflenet,BSD-3-Clause,BSD-3-Clause,TRUE,PUBLIC_QUANTIZED_VERIFIED,https://media.githubusercontent.com/media/onnx/models/4f43949841cb55a0b98dc8fcd045431ccafd9f96/validated/vision/classification/shufflenet/model/shufflenet-v2-12-int8.onnx,research/downloads/vision/VC13/shufflenet-v2-12-int8.onnx,ONNX QOperator INT8,4f43949841cb55a0b98dc8fcd045431ccafd9f96; opset 12,bf5eefde67941aa7eaa7787316c9c52c4b04e07fc9587a6ea7028883be9a8d4e,BSD-3-Clause (model-card license),upstream static PTQ via Intel Neural Compressor/ONNX Runtime,per-channel weights / per-tensor activations (scale initializer inspection),QLinearConv/QLinearMatMul internal graph; float32 input/output boundaries,int8,uint8/int8 internal; float32 boundaries,float32,float32,FALSE,PASS: ONNX 1.18 checker(full_check)+ORT 1.22.1 zero-input invoke; research/evidence/vision/runtime_results/VC13.json; FP32 research/evidence/vision/runtime_results/VC13_fp32.json,PASS: ONNX 1.18 checker(full_check)+ORT 1.22.1 zero-input invoke; research/evidence/vision/runtime_results/VC13.json; FP32 research/evidence/vision/runtime_results/VC13_fp32.json,PASS: ONNX 1.18 checker(full_check)+ORT 1.22.1 zero-input invoke; research/evidence/vision/runtime_results/VC13.json; FP32 research/evidence/vision/runtime_results/VC13_fp32.json,TRUE,https://media.githubusercontent.com/media/onnx/models/4f43949841cb55a0b98dc8fcd045431ccafd9f96/validated/vision/classification/shufflenet/model/shufflenet-v2-12.onnx,research/downloads/vision/VC13/shufflenet-v2-12.onnx,ONNX,4f43949841cb55a0b98dc8fcd045431ccafd9f96; opset 12,ea69821b4dd374ae2f33f9710dd1229ac263d0ee5b5a46ca3521f6483e1ba035,BSD-3-Clause,VERIFIED,ImageNet-1K,ImageNet terms; dataset not redistributed,Top-1 accuracy,"[1,3,224,224] NCHW","[1,1000]",RGB 224x224; ImageNet mean/std per model card inference notebook reference,Task-specific postprocessing is defined by the corresponding pipeline config,ImageNet-1K 1000 categories in the ONNX Model Zoo synset order,FALSE,UNKNOWN,FALSE,TRUE,UNKNOWN,TRUE,ONNX -> ONNX Dialect; lower stages attempted independently,research/downloads/vision/VC13/shufflenet-v2-12-int8.onnx,research/downloads/vision/VC13/shufflenet-v2-12.onnx,https://media.githubusercontent.com/media/onnx/models/4f43949841cb55a0b98dc8fcd045431ccafd9f96/validated/vision/classification/shufflenet/model/shufflenet-v2-12-int8.onnx,research/evidence/vision/sources/onnx_models/shufflenet_README.md | research/evidence/vision/runtime_results/VC13.json | research/evidence/vision/runtime_results/VC13_fp32.json,curl -fL 'https://media.githubusercontent.com/media/onnx/models/4f43949841cb55a0b98dc8fcd045431ccafd9f96/validated/vision/classification/shufflenet/model/shufflenet-v2-12-int8.onnx',200,ELIGIBLE,P1,Verified active FP32/public-quantized pair,Canonical registry reconstructed from the preserved artifact matrix and checksum-pinned model inventory after reproducibility-code recovery.