Qwen3-Coder 480B-A35B Instruct

Qwen3-Coder-480B-A35B-Instruct is a specialized causal language model built on a Mixture-of-Experts architecture, designed to serve as a highly capable assistant for software development. By utilizing 160 experts with 8 activated during each inference, the model achieves a balance between massive scale and computational efficiency. It is engineered specifically for agentic coding workflows, featuring a dedicated function call format that allows it to interact autonomously with developer environments and tools. This design intent focuses on enabling the model to handle multi-step coding tasks and generate functional applications, positioning it as a competitive alternative to proprietary models in complex programming scenarios.

The model underwent comprehensive pre-training and post-training stages to refine its reasoning and coding capabilities, resulting in state-of-the-art performance on benchmarks like SWE-bench-Verified. Its architecture, which includes 62 layers and 96 attention heads, supports a native context window of 256,000 tokens that can be extended up to 1 million tokens using Yarn, making it particularly effective for understanding large codebases. By focusing on agentic browser-use and foundational coding tasks, the model provides a robust foundation for developers seeking to automate intricate engineering workflows. Its ability to perform at a high level while maintaining an open-weight structure makes it a significant advancement for collaborative and autonomous software engineering.