Qwen AgentWorld 35B A3B

Qwen-AgentWorld-35B-A3B is a native language world model for text-generation that simulates agentic environments across seven unified interaction domains via long chain-of-thought reasoning: MCP (tool calling), Search, Terminal, SWE (software engineering), Android, Web, and OS — spanning both text and GUI environments.

The model is trained through a three-stage pipeline: continual pre-training (CPT) injects environment knowledge from more than 10 million real-world interaction trajectories; supervised fine-tuning (SFT) activates next-state-prediction reasoning; and reinforcement learning (RL with GSPO using hybrid rubric-and-rule rewards) sharpens simulation fidelity. Environment modeling is the training objective from the CPT stage onward, making this a native world model rather than a post-hoc adaptation of a general-purpose LLM.

The architecture is a Mixture-of-Experts causal language model with 35B total parameters and 3B activated parameters, supporting a context length of 262,144 tokens. The model is released under the Apache 2.0 license and is associated with the AgentWorldBench dataset.

Benchmark Performance

On AgentWorldBench — an open-ended evaluation constructed from real-world interactions of five frontier models on nine established benchmarks, scored on a five-dimensional rubric (format, factuality, consistency, realism, quality) normalized to a 0–100 scale — Qwen-AgentWorld-35B-A3B achieves the following domain scores:

This overall score surpasses Gemini 3.1 Pro (54.57), DeepSeek-V4-Pro (52.97), GLM-5.1 (51.31), Kimi K2.6 (53.42), MiniMax-M2.7 (46.12), and the base model Qwen3.5-35B-A3B (47.73), while remaining competitive with Claude Opus 4.8 (56.59) and GPT-5.4 (58.25) — the latter two being substantially larger or proprietary models.

Key Strengths

• Seven unified domains in a single model — no domain-specific checkpoints required.
• Zero-shot generalization to out-of-distribution environments (e.g., OpenClaw) supported by controllable perturbations and fictional-world construction.
• Scalable and controllable simulation — the model can serve as a decoupled environment simulator for agentic reinforcement learning, yielding gains that surpass real-environment training alone.
• Long context (262K tokens) enables multi-turn environment trajectory simulation.

For further details, see the technical report and the blog post.