openbmb/minicpm-v is a multimodal model project that handles non-text input such as images, video and audio together with language. Multimodal means dealing with different forms of input in one model. The repository holds MiniCPM-V, which looks at images and understands them, and MiniCPM-o, which takes video and audio and speaks in real time. The latest release of the GitHub repository openbmb/minicpm-v is 20250527.
The README sorts the models into two lines. MiniCPM-V 4.7 handles image, video and document understanding, and MiniCPM-o 4.5 is a 9B end-to-end (one continuous flow from input to output) omni-modal (speech, video and text in a single model) model that takes video and audio as input and generates speech and text together. MiniCPM-V 4.7 comes in a 35B-A3B MoE and a 1B variant, where MoE (Mixture of Experts) splits a model into several experts and computes only some of them. The repository description emphasizes use on local devices that see, hear and interact. It is written in Python, the license is Apache-2.0, and it was created on January 29, 2024. The star count is 26919.
That said, the range of judgment available here is limited. The change list and the release notes for the 20250527 release are not in the given record, so what this version changed cannot be confirmed. No performance benchmarks, adoption cases or comparison numbers between models are given either. The evaluation score fills only the momentum 5.9 out of 10, and no novelty, technology or quality scores are provided. How accurate the image understanding is, and how good the speech generation sounds, cannot be learned from this material alone. It fits better to record this as a group of models whose kind and scale are confirmed, and to judge the concrete changes and performance through the release notes and own measurements.