State of Open Models: Summer 2026 Observations
huggingface ·
The Hugging Face Hub continues to expand rapidly, though usage remains highly concentrated among a tiny fraction of repositories. Chinese laboratories have accelerated frontier development by bypassing incremental scaling and directly releasing models ranging from hundreds of billions to nearly three trillion parameters. Meanwhile, American hardware manufacturers like NVIDIA and AMD are increasingly contributing open-weight models to demonstrate chip capabilities, while U.S. software firms focus more on smaller architectures. Engagement metrics reveal a stark divide between community enthusiasm and actual deployment, as likes favor newly released frontier models while downloads consistently prioritize older, smaller, and more stable architectures. Licensing practices also diverge geographically, with Chinese developers widely distributing massive models under permissive licenses to drive ecosystem adoption and cloud revenue rather than direct weight sales. Qwen has emerged as the dominant foundational framework within the open community, generating over 150,000 derivative projects through its consistent release schedule and broad size range. Despite the headline-grabbing scale of modern architectures, sub-billion parameter models still capture the vast majority of downloads because they align with typical developer hardware constraints. Tools like llama.cpp now bridge this gap by enabling efficient local inference of trillion-parameter models, fundamentally changing how cutting-edge AI is deployed outside data centers.
huggingface