NVIDIA Cosmos-Reason1-7B
Robotics model7B-class physical AI reasoning vision-language model post-trained from Qwen2.5-VL-7B-Instruct with Physical AI SFT and RL for spatial-temporal common sense and embodied next-step decisions. [1] [4] [5]
Setup & usage →CheckpointsModel hub / homepageGetting started
- For
- Physical AI, robotics, and autonomous vehicle developers building high-level VLM planners, video critics, or dataset curation pipelines [1] [5]
- Parameters
- 8.29B total (7B-class: 7.07B LLM + 675.76M ViT + 545.00M projection/output layer) [5]
- Embodiments
- Embodiment-agnostic high-level reasoning across single-arm robots, bimanual robots, humanoids, egocentric human… [4] [5]
- Source rights
- Some licenses are unresolved. Review licenses
Source researched · Not tested by us
Sources checked · Research details
Licenses & related hardware
- License
- Apache-2.0, NVIDIA Open Model License · license not fully assessed for datasets · Details by part
Overview
NVIDIA Cosmos-Reason1 (nvidia-cosmos/cosmos-reason1, arXiv:2503.15558) is a 7B-class multimodal reasoning vision-language model (VLM) post-trained from Qwen/Qwen2.5-VL-7B-Instruct for physical common sense, video physical-plausibility criticism, and high-level embodied planning. At pinned commit b84a3fae3284c375bb0cefab7b53aeb1d5059ab6, the repository ships inference examples, cosmos-reason1-utils, FP8 quantization scripts, and post-training links—alongside an explicit banner stating that the project is no longer under active development and directing users to Cosmos 3 (NVIDIA/Cosmos). [1] [3] [4] [5]
Model card summary
- Architecture
Multimodal reasoning vision-language model following the Qwen2.5-VL-7B-Instruct architecture: a 675.76M-parameter Vision Transformer (ViT) encoder and 545.00M-parameter projector/output layer paired with a 7.07B-parameter dense Transformer LLM, post-trained via Physical AI supervised fine-tuning (SFT) and reinforcement learning (RL) to emit <think>...</think> chain-of-thought reasoning.
- Parameter count
8.29B total (7B-class: 7.07B LLM + 675.76M ViT + 545.00M projection/output layer)
Exact component breakdown from Hugging Face model card: ViT 675,759,104 + LLM 7,070,619,136 + output projection 544,997,376 = 8,291,375,616 parameters.
- Input modalities
- rgblanguage
Accepts text prompts plus RGB images (.jpg) or video streams (.mp4, recommended fps=4).
- Output modalities
- waypoints
Outputs natural-language chain-of-thought reasoning and high-level embodied next-step decisions (<think>...</think><answer>...</answer>, >= 4096 max_tokens recommended) rather than low-level robot action tensors; mapped to waypoints as the closest controlled-vocabulary high-level planning modality.
- Action space
Natural-language embodied reasoning and high-level next-step decisions (no direct continuous joint/pose action head); designed to act as a System-2 planner or video critic above a separate robot policy/VLA.
- Trained / supported embodiments
- Embodiment-agnostic high-level reasoning across single-arm robots, bimanual robots, humanoids, egocentric human demonstrations, and autonomous vehicles
- Training datasets
- 05/17/2025 Embodied Reasoning SFT (300.6 GB) and RL (2.6 GB) mixtures: RoboVQA (1.14M SFT), BridgeDataV2 (258k), AgiBot (38.9k), HoloAssist (273k), and NVIDIA AV (24.7k)
- 08/01/2025 augmented SFT data (2.6 TB): PLM-Video-Human (39k), Nexar (240k), Describe Anything (178k), ITS/Warehouse (24k), and Visual Critics (24k)
Checkpoints
| Checkpoint | Format | Size | Access | Pinned revision |
|---|---|---|---|---|
| nvidia/Cosmos-Reason1-7B | safetensors | 15.445 GB | open | 375e24000b24 |
4 safetensors shards totaling 16,584,414,544 bytes (15.445 GiB), ungated on Hugging Face; paper's 56B variant is not publicly released.
Is this for you?
- Use cases
- High-level embodied planning and chain-of-thought next-step reasoning as a System-2 planner for robot VLAs
- Video critic scoring of physical plausibility in synthetic or real-world robot/AV videos
- Automated physical AI dataset curation, dense temporal captioning, and Set-of-Mark (SoM) prompting for warehouse and ITS video analytics
- Known limitations
- Superseded by Cosmos 3: repository banner states Cosmos-Reason1 is no longer under active development and the Hugging Face collection is archived.
- Only the 7B-class checkpoint (nvidia/Cosmos-Reason1-7B) is publicly released; the 56B model described in arXiv:2503.15558 is not released on Hugging Face.
- Model card notes degraded accuracy on fast camera movements, overlapping human-object interactions, low lighting with high motion blur, and simultaneous multi-person actions.
- Inference is officially tested only on Linux with BF16 precision on NVIDIA Hopper/Blackwell/Ampere GPUs (>= 24 GB VRAM).
- Safety and limitations
- NVIDIA Open Model License permits commercial use and derivatives, but automatically terminates if safety guardrails are bypassed or disabled without a substantially similar guardrail.
- Outputs free-form natural language text (including potential toxic or inaccurate outputs); it does not output verified low-level robot joint or torque controls and requires downstream grounding and safety checks.
- Only tested in BF16 on Linux NVIDIA data-center GPUs (H100, A100, GB200); accuracy degrades under fast camera motion, low light, or complex overlapping interactions.
- Open Source Robotics has not downloaded the 15.45 GB weights, run vLLM/transformers inference, post-trained via cosmos-rl, or reproduced NVIDIA's benchmarks.
Compute & fine-tuning
- Inference compute
- Minimum VRAM: 24 GB
- Documented hardware: 1x NVIDIA GPU >= 24 GB VRAM (tested on H100, A100, GB200 with vLLM and transformers in BF16; FP8 PTQ script also provided)
- Precision:
bf16 - Edge / consumer capable: No
README states minimum 1 GPU with 24 GB memory; HF model card states tested on H100, A100, GB200 in BF16 only.
- Fine-tuning
- Supported: Yes
- Methods: Supervised Fine-Tuning (SFT) and Reinforcement Learning (RLHF/RL) via nvidia-cosmos/cosmos-rl; FP8 post-training quantization (W8A8-FP8) via scripts/quantize_fp8.py and llmcompressor
- Recipe: Upstream fine-tuning documentation
- Host computer
Minimum 1x NVIDIA GPU with >= 24 GB VRAM for BF16 inference (tested on A100, H100, and GB200).
- Operating system
- Linux
Model card explicitly states: 'Linux (We have not tested on other operating systems.)'
- Dependencies
- transformers >= 4.51.3, vLLM, and qwen-vl-utils
- uv and huggingface_hub CLI (for repository example scripts)
- llmcompressor >= 0.6.0 and vllm==0.9.2 (for optional FP8 PTQ)
- nvidia-cosmos/cosmos-rl (for optional SFT/RL post-training)
Setup & usage
Open the upstream getting-started guide [1] · Model card / documentation [1]
- Ensure a Linux host with at least 1x NVIDIA GPU (>= 24 GB VRAM) and install transformers >= 4.51.3 (or vLLM + qwen-vl-utils).
- Load nvidia/Cosmos-Reason1-7B directly from Hugging Face without needing the GitHub repo, or install uv and clone nvidia-cosmos/cosmos-reason1 for example scripts.
- Run uv run scripts/inference_sample.py or ./scripts/inference.py --prompt prompts/question.yaml --reasoning --videos assets/sample.mp4.
- Optional: run scripts/quantize_fp8.py (with vllm==0.9.2, transformers>=4.53.1, llmcompressor>=0.6.0) for W8A8-FP8 quantization or use nvidia-cosmos/cosmos-rl for SFT/RL post-training.
Upstream-reported benchmarks
Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.
| Benchmark | Upstream-reported score | Scope & caveats |
|---|---|---|
| Cosmos-Reason1 Embodied Reasoning Benchmark (HF model card, 2025-08-01 checkpoint) | 65.1% average accuracy across 6 domains: RoboVQA 87.3%, AV 70.8%, BridgeDataV2 63.7%, HoloAssist 62.7%, RoboFail (held-out) 57.2%, AgiBot 48.9%. | Upstream-reported accuracy on NVIDIA's Hugging Face model card (notes results differ from arXiv:2503.15558 Table 9 due to 08/01/2025 augmented training); OSR has not reproduced evaluations. |
Project notes
Architecture, checkpoint, and compute requirements
Cosmos-Reason1-7B retains the Qwen2.5-VL-7B-Instruct architecture, with approximately 8.29B parameters using the model card's component breakdown across its Vision Transformer (675.76M), dense LLM (7.07B), and projection/output layer (545.00M). NVIDIA post-trained the model in two stages—Physical AI supervised fine-tuning (SFT) and reinforcement learning (RL)—using embodied video-text mixtures including RoboVQA, BridgeDataV2, AgiBot, HoloAssist, NVIDIA AV, and (in the August 2025 update) PLM-Video-Human, Nexar, Describe Anything, ITS/Warehouse, and Visual Critics; the repository separately links nvidia-cosmos/cosmos-rl for users running custom SFT/RL post-training. It accepts RGB images or mp4 video (recommended fps=4) plus text prompts and outputs long chain-of-thought text (<think>...</think><answer>...</answer>, max_tokens >= 4096 recommended) rather than low-level joint or pose vectors. [1] [4] [5]
Only the 7B-class checkpoint (nvidia/Cosmos-Reason1-7B, HF revision 375e24000b24baed78f4618d3dd779e47cd96323, 4 .safetensors shards totaling 15.445 GiB, ungated) is publicly released; the 56B model discussed in arXiv:2503.15558 is not published on Hugging Face. Inference integrates natively with transformers >= 4.51.3 and vLLM, requiring at least one NVIDIA GPU with 24 GB VRAM; NVIDIA tested BF16 inference on Linux with A100, H100, and GB200 GPUs and also provides a W8A8-FP8 quantization script (scripts/quantize_fp8.py). [1] [5] [6]
Upstream-reported embodied reasoning benchmarks
On the nvidia/Cosmos-Reason1-7B model card (reflecting the August 2025 augmented checkpoint), NVIDIA reports 65.1% average accuracy on the Cosmos-Reason1 embodied reasoning benchmark across six domains: RoboVQA (87.3%), AV (70.8%), BridgeDataV2 (63.7%), HoloAssist (62.7%), held-out RoboFail (57.2%), and AgiBot (48.9%). NVIDIA also notes degraded accuracy on fast camera motion, overlapping human-object interactions, low lighting with motion blur, and simultaneous multi-person actions. These visual question-answering accuracy scores do not establish closed-loop robot task success, and Open Source Robotics has not downloaded the weights, run training, fine-tuning, or inference, measured inference latency, verified controller integration, or reproduced any benchmark scores. [5]
What we checked
Source researched only; no installation, training, fine-tuning, inference, benchmark, or physical hardware test performed. No affiliate links or reused media.
Not yet established
- Documented closed-loop control frequency (Hz)
Updates
Separate license matrix
Include Cosmos-Reason1-7B with Apache-2.0 code and commercially usable NVIDIA Open Model License weights (noting the guardrail termination clause and its dormant status following Cosmos 3).
| Layer | Availability / license |
|---|---|
| Code | available · Apache-2.0 Repository code, utilities, and example scripts are licensed under Apache License 2.0. [1] [2] [3] |
| Model weights | available · NVIDIA Open Model License Ungated 15.445 GiB safetensors weights commercially usable with derivative models allowed; includes automatic license termination if safety guardrails are bypassed without a substantially similar replacement. [1] [5] [6] |
| Training datasets | available · License not fully assessed Text annotations and partial videos released across nvidia/Cosmos-Reason1-SFT-Dataset, Cosmos-Reason1-RL-Dataset, and Cosmos-Reason1-Benchmark (several third-party video sources must be fetched separately; AV videos noted unavailable in card). [5] |
| Documentation | available · Apache-2.0 Repository README, tutorials (video critic, post-training, benchmarks), and detailed Hugging Face model card. [1] [2] [5] |
Apache-2.0 code vs. NVIDIA Open Model License weights
Repository source code, utilities, and example scripts are licensed under Apache-2.0. The nvidia/Cosmos-Reason1-7B weights are released under the NVIDIA Open Model License, which permits commercial use and derivative models without claiming ownership over outputs, but includes an automatic termination clause if users bypass or disable safety guardrails without a substantially similar replacement. Because the repository is in limited-maintenance mode and the Hugging Face collection is archived in favor of Cosmos 3, new deployments should weigh Cosmos-Reason1-7B against NVIDIA's newer Cosmos releases. [1] [2] [5] [7]
References
Each source links to the exact revision we read. “Used for” shows where this page relies on it.
- S001 Cosmos-Reason1 repository README
Cosmos 3 migration/maintenance-only banner, 24 GB GPU requirement, transformers >= 4.51.3 support, FP8 quantization script, cosmos-rl post-training link, and Apache-2.0 vs NVIDIA Open Model License split.Revision, date & where cited
NVIDIA · github.com · revision
b84a3fa· Accessed Nov 21, 2026
Used for: Overview, Model card summary, Checkpoints, Is this for you? and 4 more - S002 Cosmos-Reason1 LICENSE
Apache License 2.0 covering repository source code.Revision, date & where cited
NVIDIA · github.com · revision
b84a3fa· Accessed Nov 21, 2026
Used for: Separate license matrix - S003 GitHub repository metadata and pinned commit
Last commit 2026-06-07; releases Announcement-CES (2026-01-06) and Announcement-Cosmos-Reason2 (2025-12-20); Apache-2.0.Revision, date & where cited
nvidia-cosmos/cosmos-reason1 / GitHub · github.com · revision
b84a3fa· main / · Accessed Nov 21, 2026
Used for: Overview, What we checked, Separate license matrix - S004 Cosmos-Reason1: From Physical Common Sense To Embodied Reasoning (arXiv:2503.15558)
Describes physical common-sense and embodied-reasoning ontologies, two-stage Physical AI SFT and RL training, and Cosmos-Reason1-7B / 56B models.Revision, date & where cited
NVIDIA (arXiv) · arxiv.org · arXiv:2503.15558v3 abstract · Accessed Nov 21, 2026
Used for: Overview, Model card summary, Is this for you?, Architecture, checkpoint, and compute requirements - S005 nvidia/Cosmos-Reason1-7B Hugging Face model card
Exact parameter breakdown (8.29B total), Qwen2.5-VL-7B-Instruct base, NVIDIA Open Model License guardrail termination notice, training datasets, 65.1% average benchmark table, BF16/Linux test conditions, and technical limitations.Revision, date & where cited
NVIDIA / Hugging Face · huggingface.co · README.md · Accessed Nov 21, 2026
Used for: Overview, Model card summary, Checkpoints, Is this for you? and 6 more - S006 Hugging Face API metadata for nvidia/Cosmos-Reason1-7B
Verified ungated access, SHA 375e24000b24baed78f4618d3dd779e47cd96323, license_name nvidia-open-model-license, and 4 safetensors shards totaling 15.445 GiB.Revision, date & where cited
NVIDIA / Hugging Face · huggingface.co · HF API blobs=true 2026-11-21 · Accessed Nov 21, 2026
Used for: Checkpoints, Is this for you?, Architecture, checkpoint, and compute requirements, Separate license matrix - S007 Hugging Face Cosmos-Reason1 collection metadata
Collection description states it is archived and points users to nvidia/cosmos3.Revision, date & where cited
NVIDIA / Hugging Face · huggingface.co · HF collection API 2026-11-21 · Accessed Nov 21, 2026
Used for: Is this for you?, Separate license matrix
Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1
Change history
2026-11-21 — Initial 0.8.0-draft.1 model profile for NVIDIA Cosmos-Reason1-7B (Apache-2.0 code, NVIDIA Open Model License weights, 15.45 GiB safetensors).