cryptonaut/Qwen-AgentWorld-35B-A3B-heretic
cryptonaut/Qwen-AgentWorld-35B-A3B-heretic is a 35.1 billion parameter decensored version of the Qwen-AgentWorld-35B-A3B language world model, developed by Qwen. This model is specifically designed for agentic environment simulation, predicting next environment states given an agent's action and interaction history. It covers seven agent interaction domains, including tool calling, search, and software engineering, and features a 262,144 token context length for multi-turn environment simulation.
Loading preview...
Model Overview
cryptonaut/Qwen-AgentWorld-35B-A3B-heretic is a decensored variant of the Qwen-AgentWorld-35B-A3B, a native language world model developed by Qwen. This model is engineered to simulate agentic environments by predicting the next environment state based on an agent's actions and interaction history. It is built upon the Qwen3.5-35B-A3B-Base model and trained through a three-stage pipeline: Continual Pre-Training (CPT), Supervised Fine-Tuning (SFT), and Reinforcement Learning (RL).
Key Capabilities
- Native World Model: Environment modeling is a core training objective from the CPT stage, not a post-hoc addition.
- Seven Unified Domains: Covers a broad range of agent interaction environments including MCP (tool calling), Search, Terminal, SWE (software engineering), Android, Web, and OS, supporting both text and GUI interactions.
- Extended Context: Features a substantial context length of 262,144 tokens, recommended to be maintained at least at 128K for effective multi-turn environment simulation.
- Decensored Version: This specific model is a decensored iteration, showing a reduced refusal rate (42/100) compared to the original (57/100).
Performance Highlights
On the AgentWorldBench, Qwen-AgentWorld-35B-A3B demonstrates strong performance in agentic tasks, achieving an overall score of 56.39. It particularly excels in domains like MCP (64.79) and SWE (65.63), showcasing its ability to generalize to out-of-domain environments.
Good For
- Agent Development: Ideal for researchers and developers building and testing AI agents that require realistic environment simulation.
- Environment Simulation: Suitable for tasks involving predicting complex environment state transitions across diverse domains.
- Long-Context Agentic Reasoning: Benefits use cases requiring extensive multi-turn interactions and long chain-of-thought reasoning within simulated environments.