> ## Content Index
> Fetch the complete content index at: https://www.metatalks.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Most of America's biggest open models this year were built on top of Chinese ones
- URL: https://www.metatalks.ai/in-five-of-the-seven-months-of-2026-that-hugging-face-studied-the-biggest-open-model-an-american-lab-built-for-itself-was-far-smaller-than-chinas/
- Published: 2026-08-17T14:47:00.000Z
- Updated: 2026-08-17T14:46:59.000Z
- Author: Al
- Tags: News, AI Geopolitics, Frontier Models, #newswire

**The Chinese ceiling ran past 2 trillion parameters this year; the American one mostly stayed under 130 billion.**

Most American open-model releases above 100 billion parameters this year were not new models but work built on top of Chinese ones. That is one finding of Hugging Face's twice-yearly [State of Open Models study](https://huggingface.co/blog/state-of-open-models-summer-2026?ref=metatalks.ai), covering the first seven months of 2026.

In almost every month of the year, the biggest and best-performing open model out of a Chinese lab was larger than anything an American lab built for itself. The Chinese high-water mark ran from 754 billion parameters up to 2.78 trillion, while the American ceiling stayed under 130 billion in five of the seven months. Only two American models broke out of that low pattern at all: NVIDIA's Nemotron 3 Ultra, at 561 billion parameters in May and June, and Thinking Machines Lab's Inkling. The report says the shift toward Chinese labs it flagged in its spring edition is picking up pace.

Reaching that scale did not involve the usual climb. Labs have historically released small models first and worked up toward the frontier; several Chinese labs skipped the progression outright. Xiaomi and Meituan each shipped a model past a trillion parameters during the year, though a year earlier neither was a recognized name in open weights. Moonshot, MiniMax, Xiaomi and Z.ai release almost nothing at the small end, so a developer's first encounter with any of them is usually a model too big to run on local hardware. A lab no longer has to publish a compact version in-house to be within reach: quantization work across the community can make a large model usable within days of its arrival.

At that top end, AMD converted large models in volume and produced nothing original of its own, work the report describes as distributing and optimizing rather than building models. Counted on a wider basis, the American picture differs: bring in smaller models and embeddings, and U.S. participation in open source AI keeps expanding. That is the territory where Google, Microsoft, IBM Granite and OpenAI's earlier vision and speech models draw hundreds of millions of downloads a year.

The family the rest of the ecosystem builds on is Qwen, with 151,448 derivatives on the Hub, 2.6 times Meta's entire derivative presence and 4.7 times what the Llama repositories carry on their own. Those derivatives are downstream work by other developers, not releases Qwen shipped. A separate tally points the same way: of the 28,531 GGUF conversions of Qwen models the Hub hosts, 54 came from Qwen itself. Local inference leans the same way, at 39.6 million Qwen GGUF downloads a month against 20.8 million for Gemma and 7.5 million for Llama.

Leading on size is not the same as being used. Moonshot, which ships only at frontier scale, logged 37 million downloads across the year, while Qwen, releasing over a spread of sizes, took 2.05 billion among repositories that state a parameter count — roughly 55 times more. Within those repositories, models below 1B parameters account for 83% of downloads ever recorded and models above 100B for 1%. A different cut-off over a shorter window leaves that shape intact: among downloads logged during 2026, models above 70B take 3% of the volume.

Hugging Face sets limits on those counts: downloads, likes, derivatives and model releases each capture a separate facet of ecosystem activity, and none should be read as a gauge of how good a model is, how far it has been taken up commercially, or what share of the market it holds.