CPM-Live | OpenBMB

Tsinghua University NLP Laboratory, Face Wall Intelligence, and Zhihu jointly launched the OpenBMB open source multimodal large model series VisCPM. Evaluation shows that VisCPM achieves the best level in Chinese multimodal open source models.

VisCPM is an open-source multimodal large model series that supports bilingual multimodal dialogue capability (VisCPM Chat model) and text to image generation capability (VisCPM Paint model). VisCPM is trained on the billion parameter language model CPM Bee (10B), which integrates a visual encoder (Q-Former) and a visual decoder (Diffusion UNet) to support the input and output of visual signals. VisCPM can achieve excellent Chinese multimodal capability through pre training with only English multimodal data and generalization.