Why Alibaba’s Qwen‑2 Is Outperforming Global LLMs and What It Means for AI

On June 25, OpenAI announced the termination of API services for China. In response, Alibaba Cloud’s Bailei immediately offered a cost‑effective Chinese large‑model alternative, providing 22 million free tokens and dedicated migration assistance for existing OpenAI API users.

Qwen‑2 Tops Global Open‑Source Leaderboards

Just two days later, Alibaba’s Tongyi Qwen‑2 reclaimed the number‑one spot on the authoritative Open LLM Leaderboard, overtaking Meta’s Llama‑3 and France’s Mixtral. The achievement was highlighted by Hugging Face’s co‑founder and CEO, who announced the model’s top ranking on Twitter.

What Is the Tongyi Large Model?

“Tongyi” (meaning “understanding and meaning”) is Alibaba Cloud’s proprietary large language model. On December 22, 2023, Tongyi Qianwen became one of the first four domestic models to pass the national “Large‑Model Standard Conformity Evaluation,” meeting government‑defined criteria for generality and intelligence.

Record‑Breaking Performance of Qwen‑2‑72B

Within two hours of its release, Qwen‑2‑72B topped the HuggingFace Open LLM Leaderboard, achieving the highest global ranking for an open‑source model. In the LiveBench AI benchmark—developed by Yann LeCun, Abacus.AI, NYU, and others—Qwen‑2‑72B ranked first among open‑source models and was the sole Chinese model to win.

Comparisons with Other Chinese Models

In the Compass Arena evaluation by Shanghai AI Laboratory, Qwen‑2‑72B scored just one point behind GPT‑4o, placing it second overall and the highest‑ranked open‑source model. It outperformed closed‑source Chinese models such as Baidu’s Wenxin 4.0 and iFlytek’s Spark 3.5.

Significant Gains in Tongyi Qianwen 2.5

The currently deployed Tongyi Qianwen 2.5 version surpasses GPT‑4 Turbo across several metrics. Compared with the earlier 2.1 version, it improves understanding (+9%), logical reasoning (+16%), instruction following (+19%), and code generation (+10%). On the OpenCompass benchmark, Qianwen 2.5 matches GPT‑4 Turbo—a first for a domestic model.

Broad Impact on China’s AI Landscape

Large models are becoming as essential to the digital era as water, electricity, and oil. The rapid rise of Qwen‑2 demonstrates that building a home‑grown AI ecosystem is both feasible and strategically vital for China. Alibaba’s long‑term investment in AI, dating back over a decade, appears to be paying off.