AI Agents
agent-world-model
Snowflake's Agent World Model: synthesise 1,000 executable MCP tool-use environments for agentic RL.
★ 465 starsNOASSERTION
View guide →Qwen-AgentWorld: a language world model simulating MCP, terminal, web, and OS environments for agents.
Qwen-AgentWorld is an open-source ai agents project by QwenLM (QwenLM/Qwen-AgentWorld on GitHub). It has 1,024 GitHub stars and 97 forks, is released under the Apache-2.0 license, and was last updated on GitHub on 20 July 2026.
Decide whether you want to use the world model as an environment simulator for agent RL, or evaluate world models on AgentWorldBench.
Clone the repository for its domain system-prompt templates, evaluation code, and documentation.
Download the Qwen-AgentWorld-35B-A3B weights from Hugging Face (or ModelScope) and install SGLang or vLLM.
Launch the model server, then prompt it with a domain template to simulate an environment, or run the benchmark evaluation pipeline.
One model simulates MCP tools, search, terminal, software engineering, Android, web, and OS environments.
Environment modelling is the training objective from the first training stage onward, refined through supervised fine-tuning and RL on millions of real interaction trajectories described in the report.
Control instructions can inject targeted perturbations or build fully fictional, self-consistent worlds to expose agent weaknesses.
The companion benchmark scores predicted observations on format, factuality, consistency, realism, and quality across the seven domains, with a three-step infer, judge, and score pipeline.
Ready system prompts and judge prompts for each domain are provided in the repository's prompts folder.
Documented deployment on SGLang and vLLM exposing an OpenAI-compatible API, plus a Transformers inference example.
Get the prompt templates and evaluation code; weights are downloaded separately from Hugging Face.
git clone https://github.com/QwenLM/Qwen-AgentWorld.git
cd Qwen-AgentWorldPull the benchmark data if you plan to evaluate, and install the OpenAI client used by the eval pipeline.
huggingface-cli download Qwen/AgentWorldBench --repo-type dataset --local-dir ./AgentWorldBench
pip install openaiLaunch it with SGLang as an OpenAI-compatible endpoint on port 8000.
python -m sglang.launch_server --model-path Qwen/Qwen-AgentWorld-35B-A3B --port 8000 --tensor-parallel-size 4 --context-length 262144 --reasoning-parser qwen3It is developed by the Qwen team at Alibaba and released under the QwenLM GitHub organisation.
Yes. The weights and AgentWorldBench are Apache-2.0 licensed per the README; running the 35B model locally requires suitable GPU hardware, which is your own cost.
No key is needed to serve the model locally. The benchmark's judging step uses an LLM judge through an OpenAI-compatible API, which needs that provider's key.
A normal LLM answers questions; a language world model is trained specifically to predict environment observations in response to agent actions, so agents can practise inside it.
Clone the repo, download Qwen/Qwen-AgentWorld-35B-A3B from Hugging Face, and serve it with the documented SGLang or vLLM command.
Snowflake's Agent World Model: synthesise 1,000 executable MCP tool-use environments for agentic RL.
A Large Action Model framework for building AI web agents that browse and act for you.
Brave's official MCP server: web, news, image, video, local, and LLM-context search for your agents.
Bright Data's MCP server: web search, scraping, structured data and browser automation for agents (75 tools)
An open-source speech language model from Hume.
Beam — Reflection's open 501B MoE model built for agentic coding: 23B active parameters, a 1M-token context, Apache-2.0 weights, and Mirror, its open-source terminal harness
The most-starred other AI Agents projects in the LibreAI catalogue — the realistic shortlist if Qwen-AgentWorld is not the right fit.
A personal AI that really does things — on any OS, with your data staying yours.
An agentic skills framework & software development methodology that works.
Agent harness optimization system with skills, memory and security for Claude Code, Codex and more
The agent that grows with you — a Python framework for adaptive AI agents.
DeepSeek's plugin-based agent harness where everything is a plugin.
Free and open — grab the source, star the repo, and start creating.