Instructions to use MicroAGI-Labs/microcosmos-reasoner with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use MicroAGI-Labs/microcosmos-reasoner with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="MicroAGI-Labs/microcosmos-reasoner")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("MicroAGI-Labs/microcosmos-reasoner")
model = AutoModelForMultimodalLM.from_pretrained("MicroAGI-Labs/microcosmos-reasoner", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use MicroAGI-Labs/microcosmos-reasoner with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "MicroAGI-Labs/microcosmos-reasoner"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MicroAGI-Labs/microcosmos-reasoner",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/MicroAGI-Labs/microcosmos-reasoner

SGLang

How to use MicroAGI-Labs/microcosmos-reasoner with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "MicroAGI-Labs/microcosmos-reasoner" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MicroAGI-Labs/microcosmos-reasoner",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "MicroAGI-Labs/microcosmos-reasoner" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MicroAGI-Labs/microcosmos-reasoner",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use MicroAGI-Labs/microcosmos-reasoner with Docker Model Runner:
```
docker model run hf.co/MicroAGI-Labs/microcosmos-reasoner
```

Microcosmos Reasoner

A 32B vision-language Reasoner adapted by MicroAGI from NVIDIA Cosmos3-Super using curated human-egocentric data. It supports text responses to text, image and video inputs.

This release contains only the autoregressive reasoning tower, exported as Qwen3VLForConditionalGeneration with the Qwen3-VL-32B-Instruct backbone. Diffusion, image/video-generation and action-generation weights are not included.

Substantial gains from small, high-quality data

Adapting Cosmos3-Super with high-quality egocentric data improves spatial understanding within the broader domain of spatio-temporal reasoning. Pointing, counting, tracing and spatial-relation tasks probe this understanding through concrete outputs, including points and bounding boxes. Training uses 500 curated, diverse egocentric episodes, expanding to supervision on approximately 12k unique and diverse atomic action sequences.

Microcosmos Reasoner achieves 65.4% on the seven-benchmark suite and 59.8% on the expanded nine-benchmark suite, compared with 62.7% and 57.5% for Cosmos3-Super. The released Microcosmos-Nano RKL-MXFP8 achieves 63.8% and 61.7%, respectively.

On Micromazer’s visual primitives, Microcosmos improves from 13.5% to 41.7% over Cosmos3-Super before any maze-specific training, complementing the gains on the spatial benchmark suite.

Benchmark selection follows MolmoAct2 (arXiv:2605.02881, Table 3) ↗: benchmarks across spatial reasoning, pointing and embodied QA tasks. Benchmarks with restrictive licensing, data or code access issues are dropped. OpenEQA is dropped for low quality after human review. RoboSpatial is forked to remove conflicting coordinate frame instructions that caused severe parsing failures for Qwen3-family models; the fork is reported as microRoboSpatial.

Variants

Branch	Model	Precision
`main`	Microcosmos Reasoner, 32B	BF16
`nano-rkl-mxfp8`	Distilled Nano Reasoner, 8B, reverse-KL	MXFP8

The MXFP8 branch is a smaller distilled 8B model, not a quantized copy of the 32B model.

Training

LoRA fine-tuning on 500 curated egocentric episodes, with per-frame temporal-segment annotations and 57D pseudoactions derived from camera and hand state, expanding to supervision on approximately 12k unique and diverse atomic action sequences. The LoRA is merged into the model weights; no adapter is needed for inference.

Evaluation results

MicroAGI evaluations (August 2026). Scores are percentages; higher is better. Averages are unweighted.

Microcosmos is the 32B adapted Reasoner. Microcosmos-Nano RKL-MXFP8 is the distilled 8B reverse-KL model included in this release.

Spatial understanding

Larger models

Benchmark	Cosmos3-Super	Microcosmos	Gemma 4 26B-A4B-it	MiniMax-M3-MXFP4
BLINK	64.7	65.5	60.1	65.5
CV-Bench	89.4	89.2	81.2	86.1
ERQA	50.8	47.8	44.8	47.8
EmbSpatial	83.8	83.8	71.4	76.1
VSI-Bench	44.0	40.6	16.1	31.1
microRoboSpatial-Poi	31.1	54.1	11.5	10.7
microRoboSpatial-VQ	75.5	76.6	57.0	79.2
7-benchmark average	62.7	65.4	48.9	56.6
MindCube full	33.3	33.7	-	38.4
Where2Place	44.9	47.3	-	31.3
9-benchmark average	57.5	59.8	-	51.8

Nano and Gemma E2B models

Benchmark	Cosmos3-Nano	Microcosmos-Nano (forward-KL)	Nano MXFP8	Microcosmos-Nano RKL-MXFP8	Gemma 4 E2B-it	Gemma E2B-GZ s137
BLINK	60.7	59.7	60.9	60.4	43.0	44.2
CV-Bench	87.1	87.9	87.8	87.8	58.9	57.8
ERQA	44.0	48.0	48.0	48.0	34.3	34.0
EmbSpatial	81.3	81.4	81.6	81.8	46.9	47.4
VSI-Bench	42.0	38.8	38.5	38.7	15.9	20.1
microRoboSpatial-Poi	26.2	50.0	50.8	49.2	2.5	3.3
microRoboSpatial-VQ	76.3	79.1	77.7	80.4	42.3	34.4
7-benchmark average	59.7	63.6	63.6	63.8	34.8	34.5
MindCube full	40.6	40.9	40.8	41.7	28.3	30.1
Where2Place	37.2	65.2	66.6	66.9	6.2	6.0
9-benchmark average	55.0	61.3	61.4	61.7	30.9	30.8

Microcosmos-Nano is the forward-KL distilled model; Nano MXFP8 is its MXFP8 export. Microcosmos-Nano RKL-MXFP8 is the reverse-KL continuation included in this release. Gemma E2B-GZ s137 is the Golden Zone Reasoner LoRA merge of Gemma 4 E2B-it at step 137. - indicates an unavailable result, not zero.

The nine-benchmark average adds MindCube full (21,154 samples) and Where2Place (100 samples) to the original seven. Averages are unweighted means across benchmark rows, calculated from unrounded source metrics. These are the same MindCube and Where2Place measurements listed in the sections below, not additional evaluations.

microRoboSpatial’s normalized pointing protocol is not directly comparable to MolmoAct2’s RoboSpatial-Poi scores; VQ scoring is unchanged.

Micromazer: exploring visual primitives

Micromazer evaluates maze solving, tracing and counting on a fixed set of 192 requests, before maze-specific training. Microcosmos answers 80/192 correctly, compared with 26/192 for Cosmos3-Super served with SGLang. These results are separate from the seven- and nine-benchmark spatial-suite averages.

Model	Maze /96	Trace /48	Count /48
GPT-6 Astra	77.1% · 74	85.4% · 41	93.8% · 45
GPT-5.6-sol	51.0% · 49	79.2% · 38	87.5% · 42
GLM-5.3-Flash	51.0% · 49	60.4% · 29	93.8% · 45
MiMo-V2.6-Flash-RL	39.6% · 38	56.2% · 27	50.0% · 24
MiniMax-M3 MXFP4	37.5% · 36	41.7% · 20	54.2% · 26
Microcosmos	31.2% · 30	27.1% · 13	77.1% · 37
Cosmos3-Nano	47.9% · 46	12.5% · 6	8.3% · 4
Microcosmos-Nano RKL	32.3% · 31	10.4% · 5	8.3% · 4
Cosmos3-Super (SGLang)	14.6% · 14	18.8% · 9	6.2% · 3
Cosmos3-Super (vLLM)	12.5% · 12	25.0% · 12	4.2% · 2

Cells show accuracy and number correct.

Robotics and temporal understanding

Benchmark	Samples	Cosmos3-Super	Microcosmos	MiniMax-M3-MXFP4
EgoSchema subset	500	41.0	38.6	59.4
EgoPlan validation	3,343	28.0	28.4	28.4
EgoThink	N/A	69.3	69.7	N/A
MVBench	4,000	50.6	50.8	46.4
MindCube full	21,154	33.3	33.7	38.4
PhyX MC / open-ended	N/A	46.4 / 31.6	43.4 / 31.2	N/A
OmniSpatial test	1,533	50.2	49.8	56.3
5-benchmark average	30,530	40.6	40.3	45.8

The five-benchmark average excludes EgoThink and PhyX. PhyX cells show multiple-choice / open-ended scores. N/A in score columns means not evaluated, not zero.

Additional benchmarks

Benchmark	Cosmos3-Super	Microcosmos	MiniMax-M3-MXFP4	Metric
CountBench	87.8	89.4	92.3	Accuracy
RealWorldQA	76.9	75.9	67.2	Exact match
SALBench P3 / P3 box / P3 box image	74.2 / 79.3 / 74.4	72.5 / 77.6 / 71.5	86.4 / 92.0 / 78.3	Exact match
SALBench O3 / O3 box	6.2 / 6.9	10.7 / 10.9	43.4 / 46.7	Exact match
WorldVQA	12.5	11.9	25.0	Semantic-judge accuracy
Where2Place	44.9	47.3	31.3	Point-in-mask accuracy (100 samples)
LEMONADE	41.2	40.6	41.4	Video multiple choice (36,521 samples)
TOMATO	15.6	14.6	22.1	Video multiple choice (1,484 samples)
EgoPlan-Bench2	50.9	44.1	48.5	8-keyframe multiple choice (1,321 samples)

Protocol notes:

microRoboSpatial’s normalized pointing protocol is not directly comparable to MolmoAct2’s RoboSpatial-Poi scores; VQ scoring is unchanged.
Judge-scored metrics use DeepSeek-V4-Flash-0731. WorldVQA uses a non-canonical judge and is not directly comparable to its official leaderboard.
SALBench cells list variants in header order and use exact match. Where2Place checks whether a predicted point lies inside the reference mask.
EgoPlan-Bench2 uses eight keyframes per question. Its blind anti-repeat heuristic scores 49.3%; these scores alone do not establish planning capability.

Load with Transformers

import torch
from transformers import AutoProcessor, Qwen3VLForConditionalGeneration

MODEL = "MicroAGI-Labs/microcosmos-reasoner"
model = Qwen3VLForConditionalGeneration.from_pretrained(
    MODEL, revision="main", torch_dtype=torch.bfloat16, device_map="cuda"
)
processor = AutoProcessor.from_pretrained(MODEL, revision="main")

Serve with SGLang

BF16 on NVIDIA RTX PRO 6000 Blackwell (96 GB): one GPU for the 32k-context recipe, or four GPUs on one node for more context and concurrency. No quantization or CPU weight offload is required.

Requires Docker, NVIDIA Container Toolkit, a CUDA-13-compatible driver (R580+), and model access. Set HF_TOKEN locally. Commands bind the API to localhost.

export IMAGE='lmsysorg/sglang:v0.5.21-cu130@sha256:2dbe4c7f53230b09aaa8fa41891c13c51d9d2e253f486a70822aaf08783fff75'
export MODEL_REVISION='main'

One GPU — 32k context

docker run --rm --name microcosmos-tp1 --gpus '"device=0"' --shm-size=16g -p 127.0.0.1:8000:8000 -e HF_TOKEN -e TORCH_CUDA_ARCH_LIST=12.0 -e SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION=false -v "$HOME/.cache/huggingface:/root/.cache/huggingface" --entrypoint sglang "$IMAGE" serve --model-path MicroAGI-Labs/microcosmos-reasoner --revision "$MODEL_REVISION" --served-model-name microcosmos --model-type llm --tp-size 1 --dtype bfloat16 --host 0.0.0.0 --port 8000 --context-length 32768 --mem-fraction-static 0.85 --max-running-requests 1

Four GPUs — 262k context

docker run --rm --name microcosmos-tp4 --gpus '"device=0,1,2,3"' --shm-size=32g -p 127.0.0.1:8000:8000 -e HF_TOKEN -e TORCH_CUDA_ARCH_LIST=12.0 -e NCCL_P2P_LEVEL=SYS -e SGLANG_ENABLE_HEALTH_ENDPOINT_GENERATION=false -v "$HOME/.cache/huggingface:/root/.cache/huggingface" --entrypoint sglang "$IMAGE" serve --model-path MicroAGI-Labs/microcosmos-reasoner --revision "$MODEL_REVISION" --served-model-name microcosmos --model-type llm --tp-size 4 --dtype bfloat16 --host 0.0.0.0 --port 8000 --context-length 262144 --mem-fraction-static 0.85 --max-running-requests 4

Run one variant at a time and adjust GPU IDs if needed.

API example

curl --fail http://127.0.0.1:8000/v1/chat/completions -H 'Content-Type: application/json' -d '{"model":"microcosmos","messages":[{"role":"user","content":"Reply with exactly READY."}],"temperature":0,"max_tokens":512}'

Do not expose the unauthenticated API publicly. These commands serve the reasoning tower, not image/video generation.

Limitations

Results are task- and protocol-dependent: spatial gains do not imply uniformly better temporal reasoning or calibrated robot rewards. This is not a physical simulator or a safety-certified robot controller. Outputs may include reasoning text and require validation for downstream use.

License and attribution

OpenMDW License Agreement, version 1.1. Based on NVIDIA Cosmos3-Super, adapted by MicroAGI. See NOTICE for origins.

Citation

If you use Microcosmos Reasoner or its distilled Nano variant, please cite this model release:

Alex Korolev and MicroAGI Research. Microcosmos Reasoner: Egocentric Adaptation and Distillation for Spatio-Temporal Understanding. Hugging Face, 2026. doi:10.57967/hf/11100.

@misc{korolev2026microcosmos,
  author = {Korolev, Alex and {MicroAGI Research}},
  title = {Microcosmos Reasoner: Egocentric Adaptation and Distillation for Spatio-Temporal Understanding},
  year = {2026},
  publisher = {Hugging Face},
  doi = {10.57967/hf/11100},
  url = {https://huggingface.co/MicroAGI-Labs/microcosmos-reasoner}
}

Downloads last month: 24

Safetensors

Model size

33B params

Tensor type

BF16

Model tree for MicroAGI-Labs/microcosmos-reasoner

Base model

nvidia/Cosmos3-Super

Finetuned

(12)

this model

Paper for MicroAGI-Labs/microcosmos-reasoner

MolmoAct2: Action Reasoning Models for Real-world Deployment

Paper • 2605.02881 • Published May 4 • 267