Qwen releases AgentWorld-35B-A3B, a 3B-active MoE world model

View on Reddit

The model has 35B total parameters with ~3B active per token, trained to simulate MCP, terminal, SWE, Android, web, and OS environments. It is designed as a language world model that predicts environment responses rather than as a standard chat model.