Qwen3-Coder 480B-A35B Instruct

Qwen3-Coder 480B-A35B Instruct is positioned as the Qwen team's most agentic code model to date, emphasizing tool-using behavior and long-horizon coding tasks. It carries a very large 480-billion-parameter PyTorch checkpoint with no quantization applied in its reference deployment, and it is distributed as an open-weight model on both Hugging Face and ModelScope under the Qwen namespace, giving developers direct access to weights rather than only a hosted API. The "A35B" suffix in its identifier is documented in third-party catalogs but is not explained in the available excerpts, so its exact operational meaning should be treated as unverified outside of what the naming convention implies.

In practical terms, this release is aimed at programming assistants, repository-scale analysis, and agent workflows that need to call external tools and operate over very long inputs. The model is documented with English and Chinese language coverage and exposes chat plus tool-calling abilities, which together support a workflow of reading large codebases, producing patches, and invoking utilities. It is served through PyTorch-compatible inference engines such as vLLM and Transformers, and the same architecture is exposed across multiple hubs, which lowers the friction for self-hosted and on-premise code intelligence deployments where open weights and tool integration matter more than a managed endpoint.