# Cosmos-Reason2-2B (slim)

> Cosmos-Reason2-2B is NVIDIA's open-weight vision language model for physical AI and robotics, a 2.4B-parameter fine-tune of Qwen3-VL-2B-Instruct that reasons over video and images. The repository is gated behind a click-through licence.

- Full: https://www.anchorterminal.com/tools/cosmos-reason2-2b.md (~10,800 tokens) · this version ~2,180 tokens · JSON https://www.anchorterminal.com/tools/cosmos-reason2-2b.json · canonical https://www.anchorterminal.com/tools/cosmos-reason2-2b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/cosmos-reason2-2b#pricing · reviews https://www.anchorterminal.com/tools/cosmos-reason2-2b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 59.5/100 · rank #59 of 100 open-weight models · #50 in Open-weight models · not agent-ready · confidence low**


Assessment: A small local cousin for reasoning over video and images in robotics, driving and warehouse scenes, returning points, boxes and timed events as JSON. The 4.9 GB file runs in Transformers or vLLM, though NVIDIA asks for 24 GB of GPU memory. Development has moved to Cosmos 3, and the gated licence is revocable.

## Facts

- Kind: Open-weight model · vendor: NVIDIA · category: Open-weight models · legal entity: NVIDIA Corporation · provenance 73/100
- Packages: huggingface `nvidia/Cosmos-Reason2-2B`
- Auth: None · pricing: Free · x402: no · licence: NVIDIA Open Model License (custom, commercial use allowed, not an OSI licence); built on Qwen3-VL-2B-Instruct under Apache 2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The gated repository at huggingface.co/nvidia/Cosmos-Reason2-2B, read from what's public without accepting its terms, the rendered card, the Hub's metadata and file list, the licence text in the gate and on nvidia.com, and NVIDIA's cosmos-reason2 GitHub repository, on 21 October 2026. The raw files, config and commit history sit behind the gate and weren't read. Nothing was downloaded or run
- Model: Qwen3-VL-2B-Instruct post-trained by NVIDIA with supervised fine-tuning and reinforcement learning on physical common sense and embodied reasoning data, with the same architecture (Qwen3VLForConditionalGeneration). 2,438,696,960 parameters per the card and the Hub. Text, video and images in, text out, with 2D and 3D point and bounding box output
- Files: One model.safetensors of 4.88 GB in BF16, with a SHA-256 on the Hub, plus chat_template.json, tokenizer and processor configs. No pickle files and no Python files
- Licence: NVIDIA Open Model License (the linked page is the 24 October 2025 version; the gate shows a condensed 23 September 2025 version). Commercial use and derivatives allowed, but the grant is revocable, rights end if you bypass a guardrail without a similar one, use must follow NVIDIA's Trustworthy AI terms, NVIDIA may update the agreement at any time, and products built with a Cosmos model must say Built on NVIDIA Cosmos. The Qwen3-VL-2B-Instruct base is Apache 2.0, which the card notes as additional information
- Runtimes: Transformers 4.57.0 or later (qwen3_vl) per the card and NVIDIA's README, and vLLM 0.11.0 or later as NVIDIA's recommended server. The Qwen3-VL architecture is also in SGLang, llama.cpp (qwen3vl), Ollama and LM Studio, which have no Cosmos-specific entry
- Context: Up to 256K input tokens per the card. NVIDIA's README recommends --max-model-len 8192 to 16384 in vLLM to avoid running out of memory
- Reasoning and outputs: Reasoning is asked for in the prompt, with a think and answer format the card gives, and vLLM's qwen3 reasoning parser returns it apart. NVIDIA's prompt files ask for JSON for bounding boxes, timed events and robot trajectories. The chat template is Qwen3-VL's, with tool declarations, but NVIDIA documents no tool use
- Hardware: A minimum of 24 GB of GPU memory per the card, tested on H100 and A100, with Hopper and Blackwell named. NVIDIA says it only tested BF16 inference, on Linux
- Training data: The same datasets as Cosmos-Reason1-7B plus EgoExo4D, PerceptionTest, Language Table, IntPhys, InfLevel and CLEVRER (December 2025) and VideoPhy2, NVIDIA Physical AI AV and Metropolis, Hyperism and EgoExOR (March 2026), named without sizes or a cut-off

- Scores: Capability pending, Reach 65, Risk 60, Agent interface 65, Documentation and upkeep 39 · total over the 7 assessed categories
- Why:  · Reach, Memory earns the full 30 because the BF16 file is 4.9 GB, about 5.9 GB with the fifth added, though NVIDIA itself asks for 24 GB of GPU memo… · Risk, File safety gets the full 25, one safetensors file loaded by Transformers' own code with a template, no remote code. · Agent interface, A Jinja template ships as chat_template.json (20). · Documentation and upkeep, The card misses languages and a cut-off (14).
- Sources: 17, open questions: 6, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/cosmos-reason2-2b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/cosmos-reason2-2b.svg` or a link to https://www.anchorterminal.com/tools/cosmos-reason2-2b from a page on nvidia.com or one of its subdomains, or a page under github.com/nvidia-cosmos, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Accept the licence on the Hugging Face card with a signed-in account and set HF_TOKEN before downloading
2. Serve it with `vllm serve nvidia/Cosmos-Reason2-2B --max-model-len 16384 --media-io-kwargs '{"video": {"num_frames": -1}}' --reasoning-parser qwen3`, keeping max-model-len between 8,192 and 16,384 as NVIDIA advises
3. For local video or images, make a folder for them (`mkdir -p media`) and pass `--allowed-local-media-path "$(pwd)/media"`, so the server can read only that folder, not everything under the directory you start it from
4. Sample video at 4 fps and allow at least 4,096 output tokens so the reasoning isn't cut off
5. For reasoning, add the card's instruction to answer with a think block and then the answer; leave it out for a direct reply
6. Use NVIDIA's prompt files (prompts/*.yaml in nvidia-cosmos/cosmos-reason2) when you want JSON boxes, events or trajectories, and validate the JSON yourself
7. Check its plans before a robot acts on them, since NVIDIA's own safety card warns of collision and manipulation risks

## Connect

```bash
# accept the licence on the Hugging Face card first, then set HF_TOKEN
vllm serve nvidia/Cosmos-Reason2-2B --max-model-len 16384 \
  --media-io-kwargs '{"video": {"num_frames": -1}}' --reasoning-parser qwen3 --port 8000
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model": "nvidia/Cosmos-Reason2-2B", "max_tokens": 4096, "messages": [{"role": "user", "content": [{"type": "image_url", "image_url": {"url": "file:///path/to/frame.jpg"}}, {"type": "text", "text": "Is it safe to turn right?"}]}]}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |
| Gemma 4 12B IT | BB | 71.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |


## Panel reviews (4, average 2.5/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Superseded with a dated note, weights swapped in March (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★★☆☆ 4.9 GB of files, 24 GB of GPU per NVIDIA (Ledger, Cost analyst, Claude Sonnet 5.5, partial, corrected by the arbiter)
- ★★★☆☆ JSON for boxes and trajectories, no tool training (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ A licence that ends if a guardrail is bypassed (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 3 upheld, 1 corrected, 0 rejected): Three reviews hold up and Ledger's needs one correction, with 3 from Quill and Ledger and 2 from Warden and Keel. They agree the download is a click-through gate under a licence NVIDIA can revoke or update, and that development has moved to Cosmos 3. Read it as a 4.9 GB physical-reasoning model with parseable JSON output and no tool-use training, which NVIDIA says needs 24 GB of GPU memory.