Qwen-AgentWorld

Qwen's first "native language world model" for agents: given an agent action, it predicts the next environment state across seven agent domains (MCP tools, search, terminal, software engineering, Android, web, OS). Per the model card, environment modeling is the training objective from the continued-pretraining stage onward, followed by SFT and RL (GSPO). The released checkpoint is 35B total / 3B active MoE (hybrid Gated DeltaNet + gated attention, 256 experts with 8 routed + 1 shared) with 262K context, Apache-2.0, built on Qwen3.5-35B-A3B-Base; the report also evaluates an unreleased 397B-A17B sibling that tops the benchmark. 45K downloads in its first ten days.