Cosmos-Reason2-2B

# Fix list: Cosmos-Reason2-2B

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/cosmos-reason2-2b, the October 2026 research run, assessed 21 October 2026. Grade C, 59.5 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Cosmos-Reason2-2B: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Reach, 65 out of 100, up to 13.5 more on the total

Why it scored 65: 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 memory. Runtimes get 20 because the model keeps Qwen3-VL's architecture, which Transformers, vLLM, SGLang and llama.cpp run in released versions, while LM Studio's catalogue has only the base. Builds score 0, as NVIDIA publishes no quantised build of the 2B. The NVIDIA Open Model License allows free commercial use under a use policy (10). The gate is an automatic click-through that needs an account (5), and nothing hosts the 2B (0).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reach):

Whether an agent can run it, and where.

- 0 to 30, memory. The smallest build from the publisher or a runtime's own library that keeps 4-bit precision or better, at its file size plus a fifth for the runtime and a working context. 30 for 16 GB or less, 25 for 24 GB, 20 for 32 GB, 15 for 64 GB, 10 for what fits our reference machine (the GPU share of 128 GB of unified memory), 5 for 256 GB, 0 above.
- 0 to 25, runtimes. 5 for each of llama.cpp, Ollama, vLLM, SGLang, LM Studio and Hugging Face Transformers that runs the architecture in a released version, from the runtime's own code, library or documentation. A publisher's fork or an unreleased branch doesn't count. 25 at most.
- 0 to 15, builds from the publisher. 15 for quantised builds at two precisions or more (FP8, 4-bit, GGUF), 10 for one, 0 when only other people's quantisations exist.
- 0 to 15, terms of use. 15 for an OSI-approved licence with nothing added, 10 for a custom licence that lets anyone use it commercially for free below a threshold or under a use policy, 5 for research or non-commercial use only, 0 for no licence or unclear terms.
- 0 to 10, download. 10 when the files download with no account and no terms to accept, 5 for a click-through that needs an account, 0 when a person approves each request.
- 5, hosted too. The publisher's own API serves it, or at least two inference providers do, so an agent without the hardware can still call it.

When our reference-machine runs start, speed on that machine (tokens per second and time to first token at the build that fits) joins Reach.

## 2. Risk, 60 out of 100, up to 12.3 more on the total

Why it scored 60: File safety gets the full 25, one safetensors file loaded by Transformers' own code with a template, no remote code. Provenance gets 15 for files in NVIDIA's own organisation with hashes. The licence has a use policy and guardrail clause and NVIDIA can revoke or update it, so licence risk keeps 5 of 15. NVIDIA's PSIRT route and bulletins give the security programme 10. Agent guidance gets 5 for general cautions about guardrails and physical risk, and safety evaluation 0, with no safety results.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-risk):

What could go wrong when you run it, and what the publisher has done about it. A higher score means less risk.

- 0 to 25, file safety. 25 for weights in safetensors or GGUF only, 15 when pickle files (.bin, .pt, .pth) sit beside them, 0 for pickle only. 10 off when the documented way to load it or to build its prompts runs code from the repository (`trust_remote_code`, or the publisher's scripts in place of a template).
- 0 to 15, provenance. 10 when the files come from the publisher's own organisation, the one its site or docs link to, and 5 for a checksum on every file.
- 0 to 20, safety evaluation. 20 when results are published and someone other than the publisher ran them or can rerun them, 10 for the publisher's own results, 0 for none.
- 0 to 15, guidance for agent use. 15 for documented prompt-injection guidance or an instruction hierarchy the model was trained to follow, 10 for a statement of what its trained refusals cover and the safeguards to add around it, 5 for general cautions, 0 for none.
- 0 to 15, licence risk. 15 to start, 5 off for each of these, a use policy or field-of-use limits beyond the law, terms the licensor can change or revoke, a revenue or user threshold, and a licence that changed without a dated note or whose link is broken. Not below 0.
- 0 to 10, a security programme. 5 for a way to report a flaw (security.txt, a disclosure policy or a security contact), 5 for advisories handled in public.

When our reference-machine runs start, measured robustness (prompt injection with AgentDojo, over-refusal with XSTest) joins Risk.

## 3. Documentation and upkeep, 39 out of 100, up to 9.4 more on the total

Why it scored 39: The card misses languages and a cut-off (14). Training data is named by dataset without sizes or a cut-off (5). The repository last changed on 30 April 2026, 174 days ago (10). None of its 3 discussions has an NVIDIA reply (0). The 2B's weights were replaced in place in March with a dated line, and the family is now superseded by Cosmos 3 with a dated note (5). The gated files couldn't be checked against the card, which itself mixes float16 and BF16 and links a newer licence than the gate shows (5).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-upkeep):

- 0 to 20, the model card. Intended use, limitations, architecture, parameters, context length, languages and knowledge cut-off. 20 for all seven, 3 off for each that's missing.
- 0 to 20, training data. 20 for the data published or listed with sources and sizes, 15 for a data card that names the sources or the mix, the cut-off and the filtering, 10 for a general description with a cut-off, 5 without one, 0 for none.
- 0 to 20, the last change to the files, the template or the reference code. 20 within 30 days, 15 within 90, 10 within 180, 5 within a year, 0 older.
- 0 to 20, responsiveness. The publisher's replies to the 40 most recent questions and reports in the repository's discussions or issues. 20 for most, 10 for some, 5 for a few, 0 for none.
- 0 to 10, version policy. 10 for earlier versions kept in their own repositories with a dated note when one is replaced, 5 for earlier versions kept without notes or a first release, 0 when files were replaced in place without a note.
- 0 to 10, file health. 10 when config, tokenizer, template and card agree and no report of a broken file waits without a reply, 5 when one does, 0 when the files contradict the card.

## 4. Agent interface, 65 out of 100, up to 5.4 more on the total

Why it scored 65: A Jinja template ships as chat_template.json (20). Tools are declared only by the template it inherits from Qwen3-VL, with no NVIDIA documentation or parser (10), and nothing shows the 2B was trained for tool use (0). NVIDIA's prompt files ask for JSON with named keys for boxes, events and trajectories (15). Reasoning is asked for in the prompt and vLLM's qwen3 parser returns it apart (10). The card states 256K input tokens (10).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-interface):

What a model and a developer need to talk to it. Speech, image and other models with no chat interface read the tool-calling and structured-output lines for what the model takes and returns (voices, timestamps, streaming, typed outputs), and the note says how.

- 0 to 20, the chat template. 20 for a template in the repository (Jinja, in the tokenizer config or its own file), 10 when the prompt format is only given as code you run.
- 0 to 25, the tool-calling format. 25 when the publisher documents how tools are declared and how calls come back, with the parser to use in a named runtime, 10 when only the template or a runtime's own recipe says it.
- 0 to 20, tool calling the model was trained for. 20 with several calls in one turn, 10 for one call at a time.
- 0 to 15, structured output. 15 when JSON or JSON Schema output is documented for the model with how to ask for it, 5 when the card claims it without saying how.
- 10, reasoning control. Thinking that can be turned off or budgeted per request, with where the reasoning text comes back.
- 0 to 10, context. 10 for 128K tokens or more trained, 5 for 32K to 127K. Context reached only by scaling the position encoding counts one band down.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- Gated, so the raw config, tokenizer config, chat_template.json and Hugging Face commit history weren't read beyond what the Hub's public metadata shows; file health and last change rest on that
- The gate's licence text is a condensed version dated 23 September 2025, while the card links the 24 October 2025 version; whether that counts as a licence change without a note, a further 5 off licence risk, is the founder's call
- NVIDIA has moved development to Cosmos 3 (README note of 1 June 2026, collection archived). Cosmos 3 isn't listed, so supersededBy is empty
- Trained tool calling is read as 0 because NVIDIA documents no tool use; a reader who credits the inherited Qwen3-VL template would give 20
- GitHub issues on nvidia-cosmos/cosmos-reason2 weren't counted (no GitHub API in this session)
- Runtimes are credited by architecture for llama.cpp, vLLM, SGLang and Transformers, since none of the six has a Cosmos-specific entry. LM Studio's qwen3-vl-2b is the base's build, and a runtime counts only for a build of this model, so it isn't credited

## Weaknesses

- Gated behind a click-through licence that needs a Hugging Face account
- The licence is revocable, can be updated by NVIDIA and ends if a guardrail is bypassed
- NVIDIA says the repository is no longer under active development and points to Cosmos 3
- No quantised build from NVIDIA, no hosted endpoint for the 2B, and no NVIDIA reply in its discussions
- No documented tool calling, no safety results and no training data sizes or cut-off

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Accept the licence on the Hugging Face card with a signed-in account and set HF_TOKEN before downloading
- 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
- 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
- Sample video at 4 fps and allow at least 4,096 output tokens so the reasoning isn't cut off
- For reasoning, add the card's instruction to answer with a think block and then the answer; leave it out for a direct reply
- 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
- Check its plans before a robot acts on them, since NVIDIA's own safety card warns of collision and manipulation risks

## What the review panel asked for

- archive replaced weights
- memory figures by input type
- state whether tool calls are supported
- publish safety results
- a fixed licence version

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.