AI Signal 298
Chinese open models reach 2.78 trillion parameters as AMD and NVIDIA lead US release volume
↗
Hugging Face's summer 2026 open-models report finds Chinese labs setting the parameter-size frontier while US open-source publishing has shifted to hardware vendors and conversions of those Chinese models.
Open-model strategy has split into a Chinese frontier-size game and a US hardware-distribution game, which changes what an engineer actually finds when they go to download a state-of-the-art model. The report also quantifies a stark long tail, 1.5% of repositories account for 99.2% of downloads, and shows that download attention and likes attention barely overlap, so frontier releases are not the same as the models people actually use. For practitioners the practical map is: large open weights come from Chinese labs and need hardware-stack optimisation, small and embedding models still dominate usage, and most US frontier-scale open releases are derivative work.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
Chinese labs' monthly frontier ceiling ranged from 754B to 2.78T parameters in 2026, with the US ceiling staying under 130B in five of seven months; only NVIDIA's Nemotron 3 Ultra (561B, May, June) and Thinking Machines' Inkling cleared that bar.
AMD and NVIDIA each released more than 200 new open model repositories this year, and most US releases above 100B parameters are conversions or optimisations of Chinese models rather than original architectures; Meta is moving toward closed flagship models.
1.5% of Hugging Face repositories account for 99.2% of all downloads, and the top 25 most-downloaded repositories of 2026 share exactly one entry with the top 25 most-liked, none of the download top 25 were published in 2026, and 13 of 25 date from 2022.
THE CLUSTER