Chinese AI models close gap with US rivals, reshaping global race

Alibaba ↗ 0.10% Cloud's Qwen team released QVQ-72B-Preview, an experimental open-weight model that combines visual understanding with chain-of-thought reasoning, marking an early entry in what has since become a crowded field of multimodal AI systems.

A Vision Model That Thinks Out Loud

QVQ-72B-Preview, released in December 2024, builds on the Qwen2-VL-72B vision-language model and the QwQ reasoning architecture to create a system that can analyze images through step-by-step logical reasoning. The 72-billion-parameter model scored 70.3% on the Multimodal Massive Multi-task Understanding benchmark and posted strong results on MathVision and OlympiadBench, a competition-level bilingual science benchmark.simonwillison+2

The model was released under a Qwen license on Hugging Face and other platforms, allowing developers to download and run it locally. At launch, it supported only single-round dialogues with image inputs and did not handle video.huggingface+1

The Qwen team acknowledged limitations, noting that during multi-step visual reasoning, the model could gradually lose focus on image content and produce hallucinations.alibabacloud+1

Part of a Broader Open-Weight Push

QVQ-72B-Preview arrived as one piece of Alibaba's broader open-source AI strategy, which has since accelerated considerably. By March 2026, the Qwen model family had accumulated 942 million total downloads on Hugging Face, more than double the combined total of its next eight competitors. Alibaba has continued releasing increasingly capable models, including the Qwen3.8-Max flagship with 2.4 trillion parameters in July 2026 and Qwen3.8-27B in August 2026.etnownews+2

The Strategic Stakes of Open Weights

The release exemplified the divergent strategies now defining the global AI race. While American labs like OpenAI and Anthropic have concentrated on closed, API-gated models, Chinese labs including Alibaba have leaned into open-weight releases that anyone can download, modify, and deploy. As one analysis noted, safety frameworks that favor centralized API control could inadvertently disadvantage open-source alternatives — creating what some observers describe as a regulatory moat for closed-model providers. In that context, each open-weight release from Chinese labs represents not just a technical milestone but a strategic positioning in the contest over how AI is distributed worldwide.nextbigwhat+1