Qwen3-Coder is the code version of Qwen3, the large language model series developed by Qwen team.
Qwen3-Coder is the code-focused version of the Qwen3 large language model series, built by the Qwen team for agentic coding tasks. It is released in multiple sizes, including Qwen3-Coder-480B-A35B-Instruct, Qwen3-Coder-30B-A3B-Instruct and Qwen3-Coder-Next, an open-weight model designed specifically for coding agents and local development. It is aimed at developers building coding agents and tools that need strong code generation.
A developer loads the model and its tokenizer with Hugging Face transformers, formats a prompt using the tokenizer's chat template, and calls generate to produce a response; Qwen3-Coder-Next is built on the Qwen3-Next-80B-A3B-Base architecture with hybrid attention and a mixture-of-experts design. The model supports only non-thinking mode and does not emit separate reasoning blocks.
Qwen3-Coder supports a native context of 256K tokens, extendable to 1M tokens with Yarn, and understands 358 programming and markup languages. It supports fill-in-the-middle code completion in every version, works with agent platforms such as Qwen Code, CLINE and Claude Code through a specially designed function-call format, and retains strong math and general capabilities from its base model.
Qwen3-Coder's checkpoints are distributed through Hugging Face and ModelScope under names starting with Qwen3-Coder-. Function calling requires the project's updated tool parser in SGLang or vLLM, along with the matching new tokenizer and token IDs, to stay consistent with the rest of the Qwen3 series.