Beijing – Alibaba Group has officially launched Qwen3-Coder, its latest open-source artificial intelligence model designed for advanced software development. The new model, described by Alibaba as its most capable coding tool to date, is engineered to excel in complex "agentic" coding tasks. Early reactions from the developer community have been highly positive, with one user, @xjdr, stating on social media, "> qwen3_coder slaps. congrats to the qwen team."
Qwen3-Coder leverages a Mixture-of-Experts (MoE) architecture, featuring a substantial 480 billion total parameters while activating 35 billion parameters per inference, ensuring both power and efficiency. This model supports an impressive context length of 256,000 tokens natively, extendable to 1 million tokens with extrapolation methods, crucial for understanding large codebases. Its design emphasizes robust tool calling capabilities and multi-turn interaction for real-world problem-solving.
Performance benchmarks indicate Qwen3-Coder sets new state-of-the-art results among open models in agentic coding, browser-use, and tool-use tasks. On the challenging SWE-Bench Verified benchmark, the model achieved a 69.6% verified accuracy in a multi-turn interactive setting and 67.0% in single-shot mode. This places it competitively with top-tier models like Claude Sonnet 4, which recorded 70.4% accuracy, and significantly ahead of Mistral-small-2507 and GPT-4.1.
The release underscores Alibaba's strategic push in the intensifying global AI development race. Alongside the Qwen3-Coder model, the company has also open-sourced Qwen Code, a command-line interface tool adapted from Gemini Code, designed to fully unleash the model's agentic coding capabilities. This initiative aims to foster innovation within the open-source community and advance autonomous software development.