MiniCPM-V
Open multimodal models for images, video, documents and speech on-device
What it is
MiniCPM-V and MiniCPM-o are a family of open-weight multimodal language models that understand images, video, and documents, with the omni variant adding audio input and speech output. The latest release comes in a 1B dense model and a 35B-A3B mixture-of-experts model, with speculative decoding heads and GGUF builds for local inference. Developers building apps for phones, laptops, and edge chips use the repo's weights, inference code, and fine-tuning recipes to run vision-language features without a cloud API.