Alibaba's Qwen team released Qwen3-Coder-480B-A35B-Instruct in July 2025, a 480-billion-parameter Mixture-of-Experts model that activates only 35 billion parameters per inference. It achieves coding performance comparable to Claude Sonnet 4 and GPT-4 while being fully open-source under the Apache 2.0 license.
- The model utilizes 160 expert networks with 62 transformer layers, enabling autonomous multi-step problem-solving including debugging and documentation browsing.
- Trained on 7.5 trillion tokens using execution-based and long-horizon reinforcement learning across 20,000 parallel environments.
- Features a native 256,000-token context window extendable to 1 million tokens via YaRN.
- Achieved a 55.40% success rate on SWE-Bench and supports 358 programming languages.
The open release allows for unrestricted commercial use and self-hosting, offering engineering teams an alternative to proprietary API costs.