Qwen3 Coder

Qwen3-Coder is Alibaba's dedicated open-source model for agentic software engineering. The flagship variant — Qwen3-Coder-480B-A35B-Instruct — is a Mixture-of-Experts model with 480 billion total parameters and 35 billion active parameters per forward pass, released in July 2025.

**Built for agentic coding, not just completion**: Qwen3-Coder goes far beyond autocomplete. Trained with long-horizon reinforcement learning (Agent RL) across 20,000 parallel sandboxed environments on Alibaba Cloud, the model is designed to autonomously plan and execute real software tasks: - Write code for new features end-to-end - Debug existing code by reading error traces and iterating - Browse documentation and search the web for solutions - Use terminals, editors, and browsers as tools within multi-turn agentic loops

**Context and scale**: - 256K native context — holds entire large codebases in memory - Scales to 1M tokens with positional extrapolation - Competitive with leading closed-source models on SWE-Bench and similar agentic coding benchmarks

**Fully open**: Qwen3-Coder-480B-A35B-Instruct is open-source and freely available on Hugging Face under Apache 2.0. Smaller variants (7B, 32B) are also available for teams with compute constraints. Accessible via API through providers like NVIDIA NIM, Together AI, and others.

Whether you're building a coding agent, a CI/CD automation, or a development assistant, Qwen3-Coder provides enterprise-grade agentic coding capability without a license fee.