Alibaba Introduces Qwen3-Coder

Alibaba has launched Qwen3-Coder, its most powerful and agentic code model to date. The flagship version, Qwen3-Coder-480B-A35B-Instruct, is a 480B-parameter Mixture-of-Experts model with 35B active parameters. It supports a 256K token context natively and up to 1M tokens via extrapolation, achieving state-of-the-art results in Agentic Coding, Browser Use, and Tool Use, rivaling Claude Sonnet 4.

Alongside the model, Alibaba is open-sourcing Qwen Code, a CLI tool adapted from Gemini Code, optimized for agentic coding with Qwen3-Coder. The model integrates with top developer tools and supports deployment through various interfaces, including OpenAI-compatible APIs and Claude Code.

Alibaba Introduces Qwen3-Coder - Pretraining improvements include:

• 7.5 trillion tokens (70% code), with strong general and math reasoning.
• Context scaling for large repo-level tasks.
• Cleaned synthetic data via Qwen2.5-Coder.
   

Post-training advancements focus on

• Code RL: Reinforcement learning using diverse, executable coding tasks to boost real-world performance.
• Agent RL: Long-horizon reinforcement learning at scale for complex tasks (e.g., SWE-Bench), enabled by 20,000 parallel environments on Alibaba Cloud.
   

Developer access is available through:

• Qwen Code (npm or GitHub),
• Claude Code integrations,
• Cline, and
• Direct API access via Alibaba Cloud’s Model Studio.
   

Alibaba is continuing to expand Qwen3-Coder’s capabilities with new sizes and research into self-improving coding agents to reduce human workload in software engineering.