Cosmos-Reason2-2B vs Qwen-AgentWorld-35B-A3B

Cosmos-Reason2-2B 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).

Qwen-AgentWorld-35B-A3B 50

Chat template 20, chat_template.jinja in the repository, the same as the template in tokenizer_config.json. Tool-calling format 10, only the template says how tools are declared and how calls come back. Trained tool calling 10, the template renders several tool_call blocks in one turn, but the card presents the model as a simulator of what tools and environments return, not as a tool caller. Structured output 0, no JSON or JSON Schema output documented for the model. Reasoning control 0, the card says thinking is on by default and names the qwen3 reasoning parser, but doesn't say how to turn thinking off or budget it. Context 10, 262,144 tokens native per the card and the text config. 20 + 10 + 10 + 0 + 0 + 10 = 50.

Cosmos-Reason2-2B 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).

Qwen-AgentWorld-35B-A3B 55

Memory 10, Alibaba's only build is the BF16 safetensors, 69.32 GB, and 69.32 GB × 1.2 = 83.2 GB, above 64 GB and inside the reference machine's GPU share. Runtimes 20, Four released runtimes run the architecture, vLLM 0.33.0 (Qwen3_5MoeForConditionalGeneration), SGLang 0.5.23, Transformers 5.22.2 (qwen3_5_moe) and llama.cpp b12242 (its Qwen conversion code). Builds 0, Alibaba publishes the BF16 weights only. Terms 15, Apache 2.0, an OSI licence, in the metadata and as a LICENSE file, with no use policy or other terms added. Download 10, the repository isn't gated, so the files download with no account and no terms to accept. Hosted 0, Hugging Face lists no inference provider for the model (an empty provider mapping on 10 November 2026). 10 + 20 + 0 + 15 + 10 + 0 = 55.

Cosmos-Reason2-2B 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.

Qwen-AgentWorld-35B-A3B 55

File safety 25, 21 safetensors shards and no pickle or Python files, and config.json has no auto_map. Provenance 15, files sit in the Qwen organisation, the one Alibaba's Qwen repositories on GitHub link to per the Qwen3.6-35B-A3B listing (10), and the Hub gives a SHA-256 for every large file (5). Safety evaluation 0, no safety results on the card. Agent guidance 0, no prompt-injection guidance, instruction hierarchy, refusal statement or general cautions on the card. Licence risk 15, Apache 2.0 has no use policy, no threshold and no terms the licensor can change or revoke, and the card's licence link points at the LICENSE file in the repository. Security programme 0, no reporting route on the card. 25 + 15 + 0 + 0 + 15 + 0 = 55.