Qwen3-Coder is the open-weight code model family from Alibaba’s Qwen team, built specifically for agentic software engineering. Hosted on GitHub with more than 16,800 stars and actively updated in 2026, it pairs strong raw code generation with the tool-use and environment-interaction skills that autonomous coding agents need.
The family ships in several sizes, including the flagship Qwen3-Coder-480B-A35B-Instruct, the smaller Qwen3-Coder-30B-A3B-Instruct, and Qwen3-Coder-Next, which is built on the Qwen3-Next-80B-A3B base with a hybrid attention and Mixture-of-Experts architecture. The models were agentically trained at scale on executable task synthesis, environment interaction and reinforcement learning, so they handle repository-level work — navigating a codebase, editing multiple files and fixing real issues — rather than just single-function completion. Checkpoints are published on Hugging Face and ModelScope, and the models pair naturally with the Qwen Code command-line agent for terminal-based workflows.
The models support long repository contexts — up to 256K tokens natively, extendable further — which matters when an agent has to reason over a whole project rather than a single file. Their tool-calling and instruction-following were tuned specifically for agent harnesses, and the repository’s use-case gallery shows the models building complete websites, desktop utilities and even playable games from a single prompt. Because the weights are openly downloadable, the same model can power a cloud API deployment, a local coding assistant or an autonomous software-engineering agent, and the family has become a reference point for open agentic coding.
Qwen3-Coder is open-weight and free to download and self-host; you only pay for your own compute. Hosted API access is also available through Alibaba Cloud for teams that prefer not to run the models themselves.
Developers building coding agents, teams that want a powerful self-hosted code model, and researchers studying agentic software engineering.