State of Open Models: Summer 2026 Observations

Details

Hugging Face's biannual ecosystem analysis covering January-August 2026 shows the Hub growing from 2.43M to 2.96M public model repositories, 711K to 1M datasets, and 1.00M to 1.44M Spaces. The distribution remains extremely concentrated: 85.6% of models have fewer than 200 lifetime downloads, while 1.5% of repositories account for 99.2% of all downloads. The most striking finding is the divergence in frontier model size between Chinese and American labs. In nearly every month of 2026, Chinese labs -- Moonshot, MiniMax, Xiaomi, Z.ai, Tencent, and Alibaba Qwen -- released the largest publicly available open models, with monthly parameter ceilings ranging from 754B to 2.78 trillion. US labs' own-trained ceiling stayed below 130B in five of seven months, with NVIDIA's Nemotron 3 Ultra (561B) and Inkling from Thinking Machines Lab as the exceptions. Several Chinese labs skipped small models entirely: Xiaomi and MiniMax publish almost nothing below 70B, making community quantization the only path to accessibility. The report distinguishes two strategies: "frontier-only" (Moonshot, MiniMax, Xiaomi, Z.ai publish large models only, targeting benchmark position and API demand) versus "full spectrum" (Tencent, Alibaba Qwen cover from under 1B upward, targeting developer platform adoption). Both are characterized as rational but playing for different prizes. Hardware vendors have become the ecosystem's most prolific model publishers: AMD and NVIDIA each released more than 200 new model repositories, far ahead of any AI research lab. LiquidAI ranks third at roughly 100 repositories. The authors frame this as a chip-selling strategy -- open model optionality drives hardware adoption. Community quantization closes the loop: trillion-parameter models become runnable within days of any release.