Alibaba Cloud
Qwen-AgentWorld is positioned by Qwen as a native language world model for general agents. The public release includes Qwen-AgentWorld-35B-A3B open weights and AgentWorldBench, an evaluation benchmark spanning seven agent interaction domains: MCP, Search, Terminal, SWE, Android, Web and OS. The README describes the 35B-A3B model as a MoE language world model with 35B total parameters, 3B active parameters and a 256K context, trained from more than 10M real-world interaction trajectories through CPT, SFT and RL stages. It can be served through SGLang or vLLM with an OpenAI-compatible endpoint, used with Transformers, and evaluated with the included AgentWorldBench scripts.
Editorial verdict
Researchers and agent builders who need a simulator or benchmark for tool, terminal, SWE, Android, web and OS agent environments.
Avoid treating it as a general chat model or end-user IDE assistant; it is optimized for environment simulation and evaluation workflows.
Qwen-AgentWorld is tracked separately from Qwen Code because its public surface is a language world model plus AgentWorldBench, not a terminal coding product.
Apache-2.0 open weights and benchmark; self-hosted inference infrastructure required
Hugging Face download, ModelScope download, Self-hosted inference, Configured judge-model API billing
Commercial use should follow the current product, API, model license and billing terms.
Review prompt, file, media upload, retention and training-use terms before sensitive workloads.
Use the model to predict environment observations for MCP, terminal, SWE, Android, web, OS and search-style agent trajectories.
Run AgentWorldBench to score predicted observations across the repository's five evaluation dimensions.
Evaluate simulated RL, controllable perturbations and fictional-world construction before applying those ideas to real agent training.
Model names, quotas, release status, regional access and commercial terms can change quickly; recheck official sources before procurement or production use.
qwen
qwen-code
openclaw
qwenpaw
deer-flow
official · en · verified 2026-08-04
Confirms the 2026-06-24 release, 35B-A3B open weights, AgentWorldBench, seven domains, SGLang/vLLM deployment and Apache-2.0 license statement.
benchmark · en · verified 2026-06-25
Provides the research report for Qwen-AgentWorld language world models for general agents.
other · en · verified 2026-06-25
Hosts the official Qwen-AgentWorld model and AgentWorldBench dataset collection.
Last checked: 2026-08-04