qwen3-coder-480b-a35b-instruct

Qwen3-Coder-480B-A35B-Instruct is a coding-specialized large language model developed by Alibaba's Qwen Team, released as an open-source project under the Apache 2.0 license. The model is published on the Hugging Face Hub under the identifier Qwen/Qwen3-Coder-480B-A35B-Instruct, giving enterprises and developers direct access to the weights and integration pathways. It is positioned as the team's most agentic code model to date, with explicit chat and tool-use abilities documented in third-party model catalogs, making it well suited to interactive development environments and automated coding agents.

Architecturally, Qwen3-Coder-480B-A35B-Instruct uses a mixture-of-experts design totaling 480 billion parameters, with about 35 billion parameters activated on each query, balancing large-model capacity against per-request efficiency. It supports a native context length of 256K tokens (the cataloged API limit tokens), enabling it to ingest and reason over sizable codebases, long specification documents, and multi-file refactoring tasks in a single session. Post-training emphasizes code-focused reinforcement learning and long-horizon planning, which together target complex multi-step coding workflows such as tool-augmented programming, repository-level edits, and enterprise software engineering pipelines, where the combination of broad context and agentic behavior is most valuable.