mydrego/Qwen-AgentWorld-35B-A3B
Qwen-AgentWorld-35B-A3B is a 35.1 billion parameter language world model developed by Qwen, designed for agentic environment simulation across seven interaction domains. This native world model predicts environment states given agent actions, trained through a three-stage pipeline (CPT, SFT, RL) with environment modeling as its core objective. It features a 262,144 token context length and excels at zero-shot generalization to out-of-domain environments, making it a generalizable and scalable simulator for agent development.
Loading preview...
Overview
Qwen-AgentWorld-35B-A3B is a 35.1 billion parameter native language world model developed by Qwen, specifically designed for simulating agentic environments. Unlike general-purpose LLMs, its training objective from the Continual Pre-Training (CPT) stage onward is environment modeling, predicting the next environment state based on an agent's action and interaction history. It covers seven unified agent interaction domains, including tool calling (MCP), Search, Terminal, Software Engineering (SWE), Android, Web, and OS, supporting both text and GUI environments.
Key Capabilities
- Seven Unified Domains: A single model handles diverse interaction environments, from command-line interfaces to web browsing and software development.
- Native World Model: Environment modeling is a core training objective, not a post-hoc adaptation, leading to higher fidelity simulations.
- Generalizable & Scalable: Demonstrates zero-shot generalization to out-of-domain environments and allows for controllable perturbations and fictional-world construction.
- Agent Foundation Model: Provides an effective warm-up for multi-turn, tool-calling agentic tasks, showing performance transfer across 7 benchmarks, including 3 entirely out-of-domain.
- Extended Context: Features a substantial context length of 262,144 tokens, crucial for simulating long, multi-turn environment trajectories.
Performance Highlights
On the AgentWorldBench, Qwen-AgentWorld-35B-A3B achieves an overall score of 56.39, outperforming Qwen3.5-35B-A3B (47.73) and demonstrating competitive performance against larger models in specific domains like MCP and SWE.
Best Practices
For optimal simulation fidelity, users are recommended to use temperature=0.6, top_p=0.95, top_k=20 for inference and maintain a context length of at least 128K tokens. Domain-specific system prompts are provided to enhance simulation accuracy.