Qwen-Agent
Agent framework built on Qwen's instruction following, tool use, planning, and memory.
About
Qwen-Agent is a framework from the Alibaba Qwen team for developing LLM applications on top of Qwen's instruction-following, tool-usage, planning, and memory capabilities. It ships with example applications including a Browser Assistant, a Code Interpreter, and a customizable assistant, and it serves as the backend of Qwen Chat. The project added native API tool-call interface support (including vLLM's built-in tool-call parsing) and a Qwen3-Coder tool-call demo in 2025, plus the DeepPlanning agent evaluation benchmark and full documentation in early 2026.
It is aimed at developers building tool-using agents on the Qwen model family, with support for Qwen3.5 and multimodal tool calling (zoom-in, image search, web search) on Qwen3-VL. Installation is via pip with model access through DashScope, Ollama, or self-hosted vLLM, and the codebase is Apache-2.0 licensed with over 17k GitHub stars.
What it can do
- Function calling and tool use on Qwen models
- Planning and memory primitives for agent loops
- Example apps: Browser Assistant, Code Interpreter, Custom Assistant
- Native tool-call parsing incl. vLLM integration
- DeepPlanning agent evaluation benchmark
- Full documentation and cookbook notebooks
Setup
- pip install qwen-agent
- Configure a model endpoint (Qwen via DashScope, Ollama, or vLLM) per the docs at qwenlm.github.io/Qwen-Agent
- Run one of the example assistants or define your own tools
Prerequisites
Python 3.8+ and access to a Qwen-compatible model endpoint.
Use cases
Works with
Trust score — 9/10
Scored 9/10 (verified tier, metrics-v1). Evidence: Verified via official GitHub repo + qwenlm.github.io/Qwen-Agent docs, batch277; GitHub 17,181 stars; backend of Qwen Chat; DeepPlanning benchmark published Jan 2026; Trending score 88/100 from researcher assessment. License: Apache-2.0.
★ 17,181 GitHub stars
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.