State of Open Models: Summer 2026 Observations

The Hugging Face Hub's summer 2026 report reveals a shifting landscape in open-source AI. Chinese labs now dominate frontier-scale releases, with models ranging from 754B to 2.78 trillion parameters, while American labs lag behind, with NVIDIA's Nemotron 3 Ultra (561B) as a notable exception. Hardware vendors like AMD and NVIDIA have become the most prolific publishers of open models, using them to sell chips. Qwen has emerged as the community's base model, with 151,448 derivatives, 2.6 times Meta's footprint. Small models under 1B parameters account for 83% of all-time downloads, while trillion-parameter models reach users via llama.cpp, which now supports GGUF builds of models like Kimi-K3 at 2.8 trillion parameters. The report introduces a new agent-usage dataset, showing that coding agents like Claude Code and Codex are becoming major users of the Hub, with Claude Code holding 44.4% of agent traffic in July. Notably, the first documented autonomous agent intrusion occurred on Hugging Face's own infrastructure, and the analysis was completed using a quantized open model (GLM-5.2) after closed models refused. The report concludes that open-source AI is shifting from model labs to hardware and infrastructure companies, and that the ecosystem is increasingly driven by community-built derivatives and local inference tools.