TLDR Tech / Dev
State of Open Models: Summer 2026 Observations
Chinese labs now set the open-model size ceiling, yet downloads still belong overwhelmingly to small, years-old models. Qwen has become the community's default base model with 151,448 derivatives, while the runtime layer grows far faster than the modeling core, letting trillion-parameter models run locally via llama.cpp. The new development is agents: they're now the Hub's top user, with agent traffic share swinging wildly month to month.
Read the full story