Alibaba Qwen Releases 35B Language World Model for Agent Environment Simulation

Alibaba's Qwen team released Qwen-AgentWorld-35B-A3B, a 35 billion parameter language world model designed for agentic environment simulation. The model covers seven domains—MCP tool calling, Search, Terminal, Software Engineering, Android, Web, and OS—in a single model with a 262,144 token context window.

Architecture and Training

Qwen-AgentWorld-35B-A3B is built on Qwen3.5-35B-A3B-Base with 35 billion total parameters and 3 billion activated parameters using a Mixture of Experts (MoE) architecture. The model employs 256 experts with 8 activated experts plus 1 shared expert per layer.

According to Qwen, the model was trained through a three-stage pipeline: continual pre-training (CPT) to inject environment knowledge, supervised fine-tuning (SFT) to activate next-state-prediction reasoning, and reinforcement learning (RL) using GSPO to improve simulation fidelity. The team claims this makes it a "native world model" where environment modeling is the training objective from CPT onward, not a post-hoc adaptation.

The architecture uses 40 layers with a hidden dimension of 2048 and combines Gated DeltaNet and Gated Attention mechanisms. The model supports rotary position embeddings with dimension 64.

Benchmark Performance

On AgentWorldBench, Qwen-AgentWorld-35B-A3B achieved an overall score of 56.39 across seven domains, evaluated on five dimensions: Format, Factuality, Consistency, Realism, and Quality. This places it between GPT-4o (58.25) and Claude Opus 4.6 (57.80) in overall performance.

The model's strongest performance was in the MCP domain (64.79) and Software Engineering (65.63), while Search (36.69) and Web (49.55) showed lower scores. According to Qwen, the model demonstrates zero-shot generalization to out-of-domain environments and supports controllable perturbations.

Deployment and Usage

The model is available on Hugging Face and compatible with vLLM, SGLang, and Transformers. Qwen recommends running with a minimum context length of 128K tokens despite the 262K maximum, as the model leverages extended context for multi-turn environment simulation.

Recommended inference parameters are temperature=0.6, top_p=0.95, and top_k=20, with an output length of 32,768 tokens for most queries. The model uses a thinking mode by default (enclosed in <think>...</think> tags) to reason about environment state transitions before producing predictions.

Pricing information has not been disclosed. The model requires tensor parallelism across 4 GPUs for deployment.

What This Means

Qwen-AgentWorld represents a shift toward specialized models trained explicitly for agent simulation rather than repurposing general-purpose language models. The unified seven-domain approach in a single 35B parameter model suggests environment simulation may not require frontier-scale models, though performance still trails GPT-4o. The MoE architecture with only 3B activated parameters makes deployment more efficient than dense models of similar total size, potentially enabling faster local agent development workflows. Whether language world models become standard infrastructure for agent development will depend on whether simulated environments adequately replace real environment testing.