Hardware companies and Chinese research laboratories dominated new open-model releases during the first seven months of 2026, according to a report published by Hugging Face on August 14.
Public model repositories on the Hugging Face hub expanded from 2.43 million to 2.96 million over the period, datasets reached 1 million, and Spaces grew to 1.44 million. Downloads remained concentrated, with 1.5% of repositories accounting for 99.2% of all downloads and 85.6% of models recording fewer than 200 lifetime downloads.
Chinese laboratories regularly surpassed American peers in model scale, setting monthly parameter ceilings between 754 billion and 2.78 trillion. American model releases stayed under 130 billion parameters in five of the seven months. Exceptions included Thinking Machines Lab’s 952-billion-parameter Inkling, Nvidia’s 561-billion-parameter Nemotron 3 Ultra, and Arcee AI’s 399-billion-parameter Trinity-Large.
In the United States, hardware manufacturers published the largest volume of new open models. AMD and Nvidia each released more than 200 new model repositories, followed by LiquidAI with roughly 100. Google and Meta published far fewer new models than Nvidia, with Meta shifting focus toward closed flagship systems.
Licensing and Community Derivatives
Open licensing proved dominant among Chinese releases above 20 billion parameters, where 59% carried Apache 2.0 licenses and 22% carried MIT licenses, with none imposing non-commercial restrictions. DeepSeek and Z.ai released models ranging from 700 billion to 1.65 trillion parameters under MIT terms. By contrast, 29% of American releases in that parameter band carried Apache or MIT licenses, 41% used custom terms, and 30% listed no license.
Alibaba’s Qwen family accumulated 151,448 derivative repositories on the Hub, representing 2.6 times Meta’s total footprint and 4.7 times the Llama repositories. Developers created 180 to 210 new Qwen derivatives daily. Qwen models recorded 2,045 million downloads across repositories with declared parameter counts.
Local Inference and Autonomous Agents
Models smaller than 1 billion parameters accounted for 83% of all-time downloads on the Hub, while models over 100 billion took 1%. Repositories declaring the gguf library grew 464% following the ggml team joining Hugging Face in February, while Apple's mlx grew 148% and lerobot grew 194%. Qwen models generated 39.6 million monthly GGUF downloads, compared to 20.8 million for Gemma and 7.5 million for Llama.
Hugging Face’s newly published agent dataset showed Claude Code accounted for 44.4% of tracked coding agent traffic in July, down from 67.8% in April, while Codex grew to 20.8%. In July, an autonomous agent conducted an intrusion against Hugging Face infrastructure; after closed model guardrails refused to inspect the captured attack code, staff analyzed the script using an open, quantized GLM-5.2 model.
