Alibaba Qwen unveils Qwen3-Coder: its most agentic code model

Alibaba Qwen released Qwen3-Coder, their most agentic code model yet. The largest variant, Qwen3-Coder-480B-A35B-Instruct, is a 480B-parameter Mixture-of-Experts model with 35B active parameters, supporting 256K tokens natively and up to 1M tokens with extrapolation. It achieves state-of-the-art results among open models on Agentic Coding, Agentic Browser-Use, and Agentic Tool-Use, comparable to Claude Sonnet 4. Pre-training used 7.5T tokens (70% code ratio), natively supports 256K context (extendable to 1M via YaRN), and leveraged synthetic data from Qwen2.5-Coder. Post-training involved scaling Code RL on diverse real-world tasks and long-horizon RL (Agent RL) using a scalable system running 20,000 environments in parallel on Alibaba Cloud. Qwen3-Coder achieves SOTA on SWE-Bench Verified among open models without test-time scaling. Alibaba also open-sourced Qwen Code, a CLI tool for agentic coding (forked from Gemini Code), and supports integration with Claude Code and Cline. The model is accessible via API through Alibaba Cloud Model Studio.