Hugging Face released its biannual ecosystem analysis covering January–August 2026, finding a structural inversion in open-model frontier development. Chinese labs published frontier models ranging 754B–2.78T parameters every month, while American labs stayed below 130B for five of seven months, except NVIDIA's Nemotron 3 Ultra (561B, May–June) and Thinking Machines' Inkling (952B). The Hub's public model repositories grew from 2.43M to 2.96M, datasets from 711K to 1M.
Chinese labs adopted a 'size as intent' strategy: MiniMax, Moonshot, Xiaomi, and Z.ai publish almost nothing below 70B, betting on community quantization (4-bit, K-Quant) for accessibility, while Tencent and Alibaba cover 1B–multi-trillion-parameter ranges. U.S. open-model leadership shifted to hardware vendors: AMD and NVIDIA released 200+ new model repositories each (triple peers), positioning open models as chip sales enablers. At frontier scale, most U.S. releases above 100B are conversions or optimizations of Chinese models, not original architectures.
For practitioners, the download-vs.-likes split exposes the gap between developer excitement (frontier models) and production dependency (stable, small models). Chinese frontier labs drive 'likes'; established models from 2022 accumulate 99.2% of all Hub downloads across the year. The ecosystem signal: frontier optimization happens in Beijing and Hangzhou; infrastructure and deployment happen in Silicon Valley and China's hardware region.