Qwen3-Coder 480B-A35B Instruct

Qwen3-Coder 480B-A35B Instruct is the flagship variant of Qwen's open Qwen3-Coder family, designed around agentic software engineering rather than general chat. It is built as a mixture-of-experts causal language model with roughly 480 billion total parameters and about 35 billion activated per forward pass, a configuration that aims to deliver frontier-scale reasoning capacity while keeping inference cost closer to a much smaller dense model. The model is positioned by Qwen as their most agentic code release to date, emphasizing strong performance on agentic coding and agentic browser-use benchmarks where the open-source results are reported as comparable to Claude Sonnet.

A defining practical feature is its long-context design: the model natively supports 256K tokens and can be extended up to roughly one million tokens with YaRN-style extrapolation, making it well suited to repository-scale understanding where whole codebases, documentation, and tool histories must be reasoned over together. It ships with a specialized function-calling format tailored for agentic coding clients such as Qwen Code and Cline, allowing it to drive multi-step tool use, file edits, and browser interactions rather than only producing inline completions. The Apache-2.0 licensed open weights, paired with this agent-oriented tooling, make it a flexible foundation for teams building autonomous coding assistants, repository-aware chatbots, and developer workflows that demand both depth and transparency.