Open-Weight Frontier Has a New Center of Gravity, and It's Not American
HuggingFace's Summer 2026 data shows Chinese labs now own the open-weight frontier while U.S. hardware vendors quietly fill the void Meta and Google left.
2. Open-Weight Frontier Has a New Center of Gravity, and It's Not American
HuggingFace's biannual state-of-open-models report, published August 14, 2026, tracks the January-to-August period and finds Chinese labs running the largest open-weight models in nearly every month of 2026. China's monthly parameter ceiling ranged from 754B to 2.78 trillion. The U.S. ceiling stayed under 130B in five of seven months, with NVIDIA's Nemotron 3 Ultra at 561B in May and June and Thinking Machines Lab's Inkling as the only other exceptions. Xiaomi and Meituan both cleared a trillion parameters this year. Neither was a household name in open weights twelve months ago. Meanwhile, total public model repositories on the HuggingFace Hub grew from 2.43 million to 2.96 million, and datasets crossed one million.
The strategic read is sharper than a simple size race. AMD and NVIDIA each released more than 200 new model repositories in 2026, ranking first and second among all organizations on the Hub. Hardware vendors have replaced model labs as the primary U.S. contributors to open source, using freely available, hardware-optimized models as proof-of-performance for chip sales. Google and Meta, the companies that defined open model publishing in prior years, now rank well below NVIDIA in new releases. Meta's shift toward closed flagship models accelerates that gap. Most U.S. releases above 100B parameters this year are not original models but builds on top of Chinese foundations, a dependency that compounds over time.
The community quantization layer is the quiet enabler here. Labs like Moonshot, MiniMax, and Z.ai publish almost nothing below 70B, yet practitioners can run their models within days of release because the open-source tooling makes large models accessible without the lab shipping a small entry point. Watch whether U.S. policy responses to Chinese frontier dominance in open weights start targeting quantization infrastructure or export controls on the tooling layer, not just the model weights themselves.