Qwen3-Coder-480B-A35B-Instruct

Qwen3-Coder-480B-A35B-Instruct is Alibaba's flagship open code model built around a Mixture-of-Experts architecture designed specifically for agentic coding workflows. With 480 billion total parameters and 35 billion activated per inference across 160 experts, the model balances deep coding specialization against general reasoning and mathematics. Its architecture includes 62 layers and grouped query attention with 96 query heads and 8 key-value heads, enabling state-of-the-art performance among open models on coding benchmarks like SWE-bench-Verified while maintaining computational efficiency.

The model targets multi-step software engineering tasks that involve reading files, running tools, debugging failures, and iterating across real codebases. Developers can define custom tools and let the model dynamically invoke them during conversation or code generation, and it integrates natively with developer toolchains such as Qwen Code, Claude Code, and Cline. Its extended context window of 256K native tokens (expandable to 1 million via YaRN extrapolation) provides the working memory needed to hold substantial repositories and long conversation histories during extended development sessions. Released under an open-source license, it represents Qwen's most agentic code model to date.