Chinese AI Labs Surpass US in Open-Weight Model Downloads

For the year ending August 2025, Chinese open-weight artificial intelligence models captured 17.1% of global AI downloads, edging past the United States at 15.86%. That figure, drawn from a study by MIT and Hugging Face researchers, marks a concrete milestone in a competition that had until recently been framed almost entirely around raw benchmark performance.

The shift traces back to January 2025, when DeepSeek released its R1 reasoning model as a free, downloadable package. R1 matched the performance of leading American systems at reportedly a fraction of the development cost, compressing what had seemed like a widening capability gap almost overnight. The more durable win was political: giving away what rivals sell builds goodwill among developers fast.

According to MIT Technology Review, a cohort of Chinese labs now follows the same open-weight blueprint, including Z.ai (formerly Zhipu AI), Moonshot, MiniMax, and Alibaba's Qwen team. Each is releasing more capable models at a pace most Western observers did not anticipate a year ago.

The release cadence

The model tracker at llm-stats.com captures the tempo clearly. In the past two weeks alone, Moonshot pushed Kimi K2.6 on April 20, Alibaba shipped Qwen3.6-35B-A3B on April 16, and Zhipu released GLM-5.1 on April 7. All three carry open-source licenses. That rhythm is no longer a novelty; it is the baseline expectation developers now bring to Chinese labs.

Hugging Face data from last month shows Alibaba's model family, including the Qwen series, now has more user-generated variants than models from Google and Meta combined. Variants are a meaningful proxy for developer engagement: fine-tunes, merges, and domain-specific adaptations built directly on top of the base weights. A model nobody downloads never gets forked.

AI hype is fading as a driver of enterprise spending, and companies are shifting from exploratory pilots to actual deployment. In that environment, customizable and cheaper tools hold a structural advantage. Open weights let an engineering team adapt a model without negotiating a commercial agreement with an American API provider, and for teams experimenting with artificial intelligence on constrained budgets, that freedom compounds quickly across a project portfolio.

The access question

Not all open-weight releases carry permissive commercial licenses, and the hardware picture is complicated by US chip export controls. Those compute constraints may paradoxically reinforce the open-source strategy: if top-tier training hardware is harder to acquire, sharing model weights helps recruit a distributed developer community to handle fine-tuning and evaluation work that would otherwise require proprietary infrastructure at scale.

The competitive dynamic is pushing American incumbents in different directions. Anthropic this month announced Claude Mythos Preview, a model The Hill reports the company considers too dangerous for broad public release. Rather than distributing weights, Anthropic restricted access to a vetted consortium focused on patching security vulnerabilities under its Project Glasswing initiative. The contrast with the Chinese approach is deliberate, not incidental.

NVIDIA has its own open stake in the dynamic. As documented on blogs.nvidia.com, the company has released a suite of models under its Nemotron and Cosmos families, contributing open training frameworks and multimodal data at significant scale. Open models need open infrastructure, and NVIDIA benefits when developers build on hardware that runs those weights.

What this means for practitioners

The practical read is about friction. Closed API providers create dependency through rate limits, pricing changes, and deprecation schedules that teams cannot control. Open-weight models let practitioners pin a version, run inference on their own hardware, and iterate without asking permission from a vendor.

Evaluation overhead is the real tradeoff. More open-weight releases mean more models to benchmark before committing to a production stack. As the artificial intelligence review process inside engineering organizations grows more complex, inference infrastructure providers will play a larger role in helping teams deploy quickly without maintaining their own GPU clusters.

On downloads, Chinese labs now lead. On developer ecosystems measured by model variants, Qwen is ahead of every American lab in that metric. On raw capability, the gap with US frontier models is narrowing faster than most predicted. Whether openness translates into the platform leverage that closed ecosystems once monopolized is the question this industry will spend the next two years answering.

FAQ

What are open-weight AI models, and why do developers prefer them over closed APIs?

Open-weight models distribute the trained model parameters publicly, letting developers run, fine-tune, and modify the model on their own hardware. Closed APIs require routing all requests through the provider's servers, creating cost, latency, and vendor lock-in that open-weight deployments avoid entirely.

Which Chinese AI labs are releasing open-weight models in 2026?

The main players are Alibaba's Qwen team, Moonshot AI with its Kimi series, Zhipu AI with GLM, and MiniMax. DeepSeek's R1 release in early 2025 started the trend; the others have followed with increasingly capable open releases through this year.

How does Alibaba's Qwen stack up against Meta's Llama and Google's Gemma?

By user-generated variants on Hugging Face, Qwen now leads both. Variants measure real developer adoption and customization activity, making this one of the more honest signals of ecosystem traction available.

Do US chip export controls affect Chinese open-weight AI development?

They constrain access to the highest-end training hardware, but labs have adapted by optimizing training efficiency and distributing evaluation work across the developer community. The open-weight strategy may partly be a structural response to those hardware constraints.