2025 has been the most eventful year for open source AI

It is easy to forget that DeepSeek R1 was released only earlier this year. The clever use of Mixture-of-Experts, catalysed a group of open-source developers to either start or set up their game, and the open source AI landscape has evolved impressively ever since.

Amazing models like Alibaba´s Qwen family, Moonshot AI´s Kimi K2, Zhipu's GLM-4.5 and MiniMax's M2 followed, multimodality, and reasoning capabilities improved quickly, and only few impressive US and European open models followed. China is now the leading open AI model developer country.

source: Nathan Lambert and Florian Brand

Although most open models only pertain to open weights, the Swiss model Apertus and Allen Institute's OLMo 3, Molmo 2, and NVIDIA’s recently released Nemotron 3 family all deserve special mention for publishing much wider recipes for model training.

Moreover, open source received special mention in the AI Act, the Atom Project was announced and more research on the ecosystem was published: OSAI is ready for 2026.