China is winning the open-source AI race right now, not someday. According to Clément Delangue, CEO of Hugging Face (the world's largest hub for AI models), Chinese laboratories are dominating open-weight models and could catch up to the US frontier by the end of 2026. This assessment carries weight because Delangue's platform hosts models from both ecosystems, giving him a unique vantage point on the global AI landscape.

What Makes Chinese Models So Competitive?

On August 3, Delangue explained the core advantage: Chinese labs are building in the open while American labs increasingly work behind closed doors. "They're clearly dominating on open models right now," he said, "and I wouldn't be surprised if they start dominating at the frontier either by the end of this year or next year at the rate of progress". The irony is striking: American AI companies have more capital and computing power than anyone, yet this advantage may not be enough when work happens in isolation.

Delangue

The data tells a compelling story. Alibaba's Qwen model family alone accounted for more than 30% of all Hugging Face downloads in 2024, and by August 2025, Qwen derivatives made up over 40% of new language-model variants published on the platform, compared to roughly 15% for Meta's Llama. Add in models from other Chinese labs, and the picture becomes even clearer.

Which Chinese Models Are Leading the Pack?

Several Chinese open-weight models are now setting the pace across major benchmarks:

• GLM-5.2: Currently tops the Artificial Analysis Intelligence Index among open models, demonstrating strong performance across a range of tasks.
• Kimi K3: Became the first open model to lead the Arena.ai Frontend Code Arena, a crowdsourced benchmark for web development tasks.
• DeepSeek-V4-Pro-Max: Leads raw SWE-bench Verified scores among downloadable weights at 80.6% and ships under a permissive MIT license, making it freely available.
• Qwen3.8-Max: Alibaba's newest 2.4 trillion-parameter model, launched on August 3, performed comparable to or higher than leading US models like Anthropic's Fable 5 on multiple benchmarks.

Alibaba's Qwen3.8-Max is particularly notable because it costs significantly less than US alternatives. At $2 per million input tokens and $6 per million output tokens, it undercuts Anthropic's Fable 5, which costs $10 for input and $50 for output. Alibaba plans to release the open weights of Qwen3.8-Max the week after launch, another differentiator from leading US labs that keep their models proprietary.

Why Open Models Create a Compounding Advantage?

The real competitive edge for Chinese labs lies in what Delangue calls a "network effect." Open models compound in power because every derivative, fine-tune, and research paper built on top of them feeds back into the ecosystem and speeds up the next release. Closed labs, by contrast, only get what their own researchers produce on their own timeline, without thousands of outside contributors stress-testing and improving the work for free.

This isn't simply a compute gap or an engineering problem. It's a business-model problem. An enterprise choosing between a closed US model and an open Chinese one used to be weighing capability against control. Now, as the capability gap closes, control and cost, plus the freedom to self-host and fine-tune, are starting to win the argument outright.

How to Evaluate Chinese vs. US AI Models for Your Needs

• Benchmark Performance: Check independent, crowdsourced benchmarks like Code Arena rather than relying solely on company-published results. Qwen3.8-Max placed fourth on Code Arena for front-end web development, behind Claude Opus 5, Kimi K3 Max, and Claude Opus 5 High.
• Cost Structure: Compare token pricing across models. Chinese models typically cost 5 to 25 times less than US frontier models, making them attractive for cost-sensitive deployments.
• Open-Weight Availability: Determine whether you need the ability to self-host, fine-tune, or modify the model. Chinese labs increasingly release open weights; US labs typically do not.
• Real-World Task Performance: Benchmarks don't always reflect real-world performance. Test models on your specific use case before committing.

What Do Experts Say About the Capability Gap?

Not everyone agrees that Chinese models have caught up to the US frontier. Andrew Yoon, a technical staff member at CivAI, cautioned against overstating the achievements. "Claims that recent Chinese models match or beat the US frontier are overstating cherry-picked benchmark results," Yoon stated. "This isn't to say that the models are weak. Models like Qwen 3.8 Max, Kimi K3, and GLM 5.2 are all very capable, but they continue to lag the US frontier by a significant margin".

"Claims that recent Chinese models match or beat the US frontier are overstating cherry-picked benchmark results. This isn't to say that the models are weak. Models like Qwen 3.8 Max, Kimi K3, and GLM 5.2 are all very capable, but they continue to lag the US frontier by a significant margin," stated Andrew Yoon.

Andrew Yoon, Technical Staff Member at CivAI

However, Delangue's perspective suggests the gap is narrowing faster than many realize. The key distinction is that American frontier labs like OpenAI, Anthropic, and Google DeepMind still hold real leads in the largest, most expensive models, and closed systems still dominate the highest end of enterprise deployment. But momentum, Delangue argues, is running through Hangzhou and Beijing, not San Francisco.

What This Means for the AI Industry Going Forward

The rise of Chinese open-weight models is reshaping how enterprises think about AI deployment. Rather than choosing between capability and cost, organizations can increasingly have both. The open-source ecosystem also means that improvements compound faster, as thousands of researchers and developers contribute improvements without waiting for official releases from a single company.

Alibaba's marketing approach for Qwen3.8-Max also signals a shift in how AI is being positioned to the public. Instead of emphasizing productivity gains, the company frames the model as an "always on work-mate" that gives people back their time, showing examples of engineers going fishing while the model designs chips for 12 hours, or professors playing tennis while the model verifies proteins. This lifestyle-focused messaging contrasts sharply with how US labs have marketed AI and may resonate more with skeptical consumers concerned about job displacement.

The bottom line: Chinese labs have moved from catching up to setting the pace in open-weight AI. Whether they reach the US frontier by late 2026 remains to be seen, but the trajectory is clear, and the competitive pressure on US labs is intensifying.