qwenlm/qwen3-coder

Qwen3-Coder is the code-focused branch of the Qwen3 family, a set of large language models built by the Qwen team. A large language model is the kind of AI that powers chat assistants and code helpers, the "coder" variant is one that has been further trained to be especially good at reading and writing programming code and at acting as the brain behind agentic coding tools, software that can take a task description and then plan, edit files and run commands on its own. According to the README the project ships several model sizes, including Qwen3-Coder-480B-A35B-Instruct, Qwen3-Coder-30B-A3B-Instruct, and a smaller Qwen3-Coder-Next built on Qwen3-Next-80B-A3B-Base. Qwen3-Coder-Next is open-weight, meaning the trained model parameters can be downloaded and run by anyone rather than being available only through a paid API. The README describes a hybrid architecture combining attention and a mixture-of-experts setup, training that emphasises agentic tasks (synthesised executable tasks, environment interaction, and reinforcement learning) and a function-call format aimed at coding agents. The models support a native context window of 256K tokens, extendable to 1M, and the README lists 358 programming languages handled by the model. People would use Qwen3-Coder to power code assistants, autonomous coding agents, and tooling for repository-scale understanding where a very long context helps. The repo links to companion tools and integrations such as Qwen Code, CLINE and Claude Code, plus a WebDev demo space. The repository's primary language is Python, the model weights are hosted on Hugging Face and ModelScope, and the full README is longer than what was provided here.