Alibaba’s Qwen dominates open‑source AI downloads, while new study teaches models to code

Alibaba Cloud’s Qwen 3.5 release in February sparked a surge in downloads, with 153.6 million pulls in February alone, more than twice the combined total of the next eight major models, Interconnects AI reports, citing Hugging Face data. The same newsletter notes that Qwen’s cumulative downloads approached one billion by March, giving the Chinese suite over 50 percent of all open‑source AI model downloads worldwide.

The download dominance reflects a broader shift toward Chinese‑origin models. Interconnects AI says Chinese offerings overtook U.S. counterparts last summer and have kept the lead, even as OpenAI and Nvidia begin to make modest gains in the open‑source arena. Alibaba Cloud positions Qwen 3.5 as “on par” with leading U.S. models from OpenAI and Anthropic, the company announced in its February open‑source release.

In parallel, researchers at arXiv have unveiled a new method to teach language models to code by mimicking real student behavior. Their paper, “Teaching Language Models How to Code Like Learners,” describes a conversational serialization of student log traces, turning each code submission and test feedback into dialogue turns. The approach combines supervised fine‑tuning with preference optimization to align models with authentic debugging patterns.

The team trained Qwen‑4B and Qwen‑8B models on a large dataset of real Python assignment submissions. Results show the Qwen‑based learners outperform prior code‑only methods and prompted large‑language‑model baselines in functional alignment and code similarity, according to the arXiv preprint. The authors also released their code to enable reproducibility.

Both developments underscore Alibaba’s expanding AI footprint: a market‑leading open‑source model suite and a research pipeline that enhances Qwen’s coding capabilities using genuine student data. The dual thrust could accelerate adoption of Chinese models in education and enterprise settings, while offering a cost‑effective alternative to proprietary APIs.