Touted as the company’s “most advanced agentic AI coding model to date”, the Qwen3-Coder-480B-A35B-Instruct – built on a so-called Mixture of Experts (MoE) architecture – features a total of 480 billion parameters, 35 billion of which are active, and supports a 256,000-token context window, expandable to 1 million tokens through extrapolation methods.
In a blog entry on developer platform GitHub, the Qwen team of Alibaba Cloud – the AI and cloud computing arm of Alibaba, owner of the Post – showed that Qwen3-Coder had set new state-of-the-art results among open-source models across multiple agentic tasks, outperforming Kimi K2 Instruct from Moonshot AI and DeepSeek’s V3-0324.
On SWE-Bench, a benchmark for evaluating large language models on real-world software engineering tasks, Qwen3-Coder showed stronger capabilities than those of the two Chinese AI start-ups’ models.
The same benchmark tests also showed that Qwen3-Coder was rated either on par or just behind in certain areas when compared with the performance of prominent proprietary AI models Claude Sonnet-4 from Anthropic and OpenAI’s GPT-4.1.
“Agentic AI coding is transforming software development by enabling more autonomous, efficient and accessible programming workflows,” Alibaba said in a statement on Wednesday. “With its open-source availability, strong agentic coding capabilities, and seamless compatibility with popular developer tools and interfaces, Qwen3-Coder is positioned as a valuable tool for global developers in software development.”
Source link
