Alibaba releases Qwen2.5-Coder

Alibaba’s Qwen team released Qwen2.5-Coder, an open-weight family of code-specialised models built on the Qwen2.5 base, in six sizes from 0.5 billion to 32 billion parameters, each with a 128,000-token context window intended to handle whole code repositories rather than isolated snippets.

Alibaba reported that the largest, 32-billion-parameter Instruct variant scored competitively with GPT-4o on multiple code-generation benchmarks, including EvalPlus, LiveCodeBench and BigCodeBench, and reported the best results among open-weight coding models on several of them at release. The family supported more than 40 programming languages. As with most vendor-reported benchmark comparisons, these figures came from Alibaba’s own evaluation and were not independently reproduced at release, though independent developers who ran the 32B model subsequently reported it performing close to Alibaba’s claims on everyday coding tasks.

The release continued a rapid cycle of open-weight coding models from Chinese labs through 2024, following DeepSeek Coder and its successors, and reinforced a pattern in which Chinese open-weight releases were closing the gap with closed frontier models specifically on code — a domain where benchmark scores were relatively easy to verify against real repositories, making the competitive claims harder to dismiss than on more subjective tasks.