Qwen3-Coder Unleashed: Meet the 480B-Parameter Open Code Model Dominating Agentic Coding Benchmarks

Qwen3-Coder has arrived, and it’s setting new standards for open agentic code models. The star release, Qwen3-Coder-480B-A35B-Instruct, boasts a massive 480 billion parameters as a Mixture-of-Experts model (with 35 billion active parameters), natively supporting an impressive 256K context length—extendable to 1M via extrapolation.

Performance-wise, it’s a standout: the model clinches top-tier results across multiple agentic coding benchmarks, including the highly regarded SWE-bench-Verified, making it a go-to choice for developers seeking robust coding assistance.

But that’s not all. Alongside the model, the team is open-sourcing Qwen Code, a command-line tool built for agentic coding. Forked from Gemini Code, it comes with custom prompts and function call protocols to fully harness Qwen3-Coder’s capabilities. What’s more, Qwen3-Coder integrates smoothly with the community’s leading developer tools, ensuring seamless workflow integration.

As a foundation model, Qwen3-Coder aims to power agentic coding across the digital landscape—truly embodying "Agentic Coding in the World."

• Chat: https://chat.qwen.ai
• Blog: Qwen3-Coder: Agentic Coding in the World
• Model: Qwen3-Coder-480B-A35B-Instruct
• Qwen Code: github:qwen-code