A report analyzing Hugging Face Hub data from early 2026 reveals that Chinese labs have surpassed American ones in frontier open model size, while hardware vendors like NVIDIA and AMD dominate new repository releases. The analysis highlights a divergence between community excitement (likes) and actual adoption (downloads), with small models remaining the practical standard for local inference.
- Chinese labs released monthly ceiling models ranging from 754B to 2.78 trillion parameters, while U.S. ceilings stayed under 130B in five of seven months.
- NVIDIA and AMD each released over 200 new model repositories, far ahead of other organizations, using open models to demonstrate hardware capabilities.
- Qwen has become the ecosystem's base model with 151,448 derivatives on the Hub, significantly outpacing Meta’s Llama footprint.
- Models under 1B parameters account for 83% of all-time downloads, while those above 100B take only 1%, driven by local inference tools like llama.cpp.
- Chinese labs license large models permissively, with 59% using Apache 2.0 and 22% MIT, avoiding non-commercial restrictions to drive API and ecosystem growth.
The data suggests that open source value has shifted from model licensing revenue to hardware positioning and ecosystem dominance, with quantization layers enabling the viability of frontier-first release strategies.