cryptonaut/Qwen-AgentWorld-35B-A3B
Qwen-AgentWorld-35B-A3B by Qwen is a 35.1 billion parameter native language world model, with 3 billion activated parameters, designed for agentic environment simulation. It covers seven interaction domains including tool calling, search, and software engineering, predicting environment states through long chain-of-thought reasoning. This model excels at simulating complex agentic environments and serves as an agent foundation model, offering generalizable and controllable simulation capabilities.
Loading preview...
Qwen-AgentWorld-35B-A3B: A Native Language World Model for Agentic Simulation
Qwen-AgentWorld-35B-A3B is a 35.1 billion parameter (3B activated) causal language model developed by Qwen, specifically engineered as a native language world model. Unlike general-purpose LLMs adapted for simulation, environment modeling is its core training objective from the initial Continual Pre-Training (CPT) stage, followed by Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL).
Key Capabilities & Differentiators
- Seven Unified Domains: This single model simulates agentic environments across seven distinct domains: MCP (tool calling), Search, Terminal, SWE (software engineering), Android, Web, and OS, encompassing both text and GUI interactions.
- Native World Model Design: Its architecture and training pipeline are intrinsically focused on predicting the next environment state given an agent's action and interaction history, enabling long chain-of-thought reasoning for simulation.
- Generalizable & Controllable Simulation: The model demonstrates zero-shot generalization to out-of-distribution environments (e.g., OpenClaw) and allows for controllable perturbations and fictional-world construction, surpassing real-environment training limitations.
- Agent Foundation Model: It provides a strong foundation for agents, with LWM RL warm-up transferring effectively to multi-turn, tool-calling agentic tasks across various benchmarks.
- Extended Context Length: Features a notable context length of 262,144 tokens, crucial for multi-turn environment simulations.
Performance Highlights
On the AgentWorldBench, Qwen-AgentWorld-35B-A3B achieves an overall score of 56.39, performing competitively across domains like MCP (64.79) and SWE (65.63), demonstrating its robust simulation fidelity.
When to Use This Model
This model is ideal for developers and researchers focused on:
- Building and evaluating AI agents that require realistic and complex environment simulations.
- Developing systems that need to predict and reason about the outcomes of agent actions in diverse digital environments.
- Researching agentic AI, particularly in areas requiring controllable and scalable simulation platforms.