Alibaba Qwen Team · Code Generation

Qwen3-Coder

480B-parameter MoE model with 35B active parameters trained on 7.5T tokens, supporting 256K native context and 1M with extrapolation, achieving state-of-the-art performance on agentic coding tasks through scaled Code RL and long-horizon Agent RL training. Open-sourced with Qwen Code CLI tool and Claude Code integration for real-world software engineering tasks.