The top-100 liked open models in the first 6 months of 2025 show a significant set of DeepSeek models, as well as many new developers compared to last years.
DeepSeek-R1 was already released on January 20th. However, model versions from March and May made the list as well, and the most popular multimodal model is the 7B (Billion parameter) version of Janus-Pro.
Computer vision is turning into video generation, with Wan-AI uploading various video generating models, joining ByteDance, StepFun and Black Forest Labs, amongst others.
The list also includes many smaller models under 4B, such as versions from Qwen, JetBrains and DeepSeek (NLP), Gemma and SmolDocling (multimodal) and audio models from Nari Labs, Sesame and Invidia. Especially the rise in variety of audio models is interesting compared to last year (see below), for both Text-to-Speech and Speech-to-Text users start to be able to access more and more varieties.