Qwen3 Coder 480B A35B Instruct was introduced by the Qwen team as their most agentic code model to date, built around a Mixture-of-Experts architecture with 480 billion total parameters and 35 billion active per forward pass, routing through 8 of 160 experts. The variant is optimized for code generation, function calling, tool use, and long-context reasoning across software repositories, and the team positions it for agentic coding, agentic browser-use, and agentic tool-use workloads where it reportedly matches the strongest closed competitors on open-model benchmarks. It is published as an open-weight release under the Qwen organization, which is what allows the FP8 build to be redistributed and served by third-party hosts.
In practical terms, the model handles very large codebases through a 256K-token native context window that can be extended to roughly one million tokens via extrapolation methods, while the API surface exposed through third-party routers reports a 262K working window and supports both text input and text output with temperature control. It pairs with a dedicated command-line agent, Qwen Code, adapted from Gemini Code with custom prompts and function-calling protocols, making it a natural fit for developers who want to drive repository-scale refactors, multi-step debugging, and tool-augmented coding pipelines. The result is a long-horizon coding assistant that combines open weights, a substantial active-parameter budget, and extended context for agent-style development work.
