Overview
Qwen3-Coder is the Qwen team's model line built specifically for agentic coding: multi-step tasks where the model reads a repository, calls tools, runs commands and edits several files. The flagship, Qwen3-Coder-480B-A35B-Instruct, is a mixture-of-experts (MoE) model that activates only 35B of its 480B parameters per token, which keeps inference cost far below a dense model of the same size.
At release it was among the strongest open-weight models on agentic benchmarks such as SWE-bench Verified, competitive with leading proprietary models. The Qwen team also open-sourced Qwen Code, a command-line agent adapted to the model.
Strengths
- Trained with reinforcement learning on real software-engineering tasks, not only code completion
- 256K-token native context, enough to load large parts of a repository at once
- Strong function-calling and tool-use behavior for agent frameworks
- Permissive Apache 2.0 license
Run it locally
The smaller MoE variant is the practical choice on a workstation:
ollama run qwen3-coderFor the 480B model, serve it with vLLM or SGLang on a multi-GPU node and point your agent at the OpenAI-compatible endpoint.
Good fit for
Autonomous coding agents, repository-scale refactors and long-context code review where you want an open, commercially usable model.
Frequently asked questions
Is Qwen3-Coder free for commercial use?
Yes. Qwen3-Coder weights are released under the Apache 2.0 license, which allows commercial use, modification and redistribution.
Can I run Qwen3-Coder locally?
The 30B-A3B variant runs on a single high-memory GPU or a recent Apple Silicon Mac with quantization. The 480B flagship needs a multi-GPU server.
What context length does Qwen3-Coder support?
256K tokens natively, and up to roughly 1M tokens with YaRN extrapolation.