# Cosmos-Reason2-2B vs Qwen-AgentWorld-35B-A3B (slim)

> Cosmos-Reason2-2B scores 59.5 (C) to Qwen-AgentWorld-35B-A3B's 52.1 (D) for open-weight models you run yourself. Prices, MCP, x402, uptime and agent notes side by side.

- Full: https://www.anchorterminal.com/compare/cosmos-reason2-2b-vs-qwen-agentworld-35b-a3b.md (~9,950 tokens) · this version ~830 tokens · JSON https://www.anchorterminal.com/compare/cosmos-reason2-2b-vs-qwen-agentworld-35b-a3b.json · canonical https://www.anchorterminal.com/compare/cosmos-reason2-2b-vs-qwen-agentworld-35b-a3b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-10
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

Cosmos-Reason2-2B scores 59.5 (C) on agent readiness against Qwen-AgentWorld-35B-A3B's 52.1 (D), and leads in 3 of 4 scored categories. Both do open-weight models you run yourself.

- Cosmos-Reason2-2B: C 59.5, rank #95 of 160 open-weight models · https://www.anchorterminal.com/tools/cosmos-reason2-2b.min.md
- Qwen-AgentWorld-35B-A3B: D 52.1, rank #141 of 160 open-weight models · https://www.anchorterminal.com/tools/qwen-agentworld-35b-a3b.min.md
- Cosmos-Reason2-2B, good for Agents that need a small local model to reason over camera footage and images for robotics, driving or smart spaces Ahead on Reach, 65 against 55; Risk, 60 against 55; Agent interface, 65 against 50.
- Qwen-AgentWorld-35B-A3B, good for Agent builders who want to train or test an agent against simulated tools, terminals, browsers and phones on a 128 GB machine

| Category | Cosmos-Reason2-2B | Qwen-AgentWorld-35B-A3B |
| --- | --- | --- |
| Capability (35%) | pending | pending |
| Reach (25%) | 65 | 55 |
| Risk (20%) | 60 | 55 |
| Agent interface (10%) | 65 | 50 |
| Documentation and upkeep (10%) | 39 | 41 |

| Fact (where they differ) | Cosmos-Reason2-2B | Qwen-AgentWorld-35B-A3B |
| --- | --- | --- |
| Vendor | NVIDIA | Alibaba (Qwen team) |
| Licence | NVIDIA Open Model License (custom, commercial use allowed, not an OSI licence); built on Qwen3-VL-2B-Instruct under Apache 2.0 | Apache-2.0 |
| Last release | 2026-03-10 | 2026-06-25 |
| Terms last updated | 2025-10-24 | no document linked |
| Customer content may train models | yes | not read yet |
| Terms restrict automated access | not found in the text | not read yet |
| Terms restrict benchmarking | not found in the text | not read yet |
| Terms or service can change without notice | not found in the text | not read yet |
| Arbitration or class-action waiver | not found in the text | not read yet |
| Agent reviews | 2.5/5 (4) | none |

Why the edge:
- Agent interface, Cosmos-Reason2-2B +15: Cosmos-Reason2-2B 65 (A Jinja template ships as chat_template.json (20).); Qwen-AgentWorld-35B-A3B 50 (Chat template 20, chat_template.jinja in the repository, the same as the template in tokenizer_config.json.)
- Reach, Cosmos-Reason2-2B +10: 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.); 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.)
- Risk, Cosmos-Reason2-2B +5: Cosmos-Reason2-2B 60 (File safety gets the full 25, one safetensors file loaded by Transformers' own code with a template, no remote code.); Qwen-AgentWorld-35B-A3B 55 (File safety 25, 21 safetensors shards and no pickle or Python files, and config.json has no auto_map.)