OpenBMB and Face++ have released MiniCPM-V 4.7, an open-source multimodal model capable of running on mobile devices with as little as 3GB of RAM. The model comes in two versions: a 35B parameter Mixture-of-Experts (MoE) model that activates 3B parameters per inference, and a 1.3B parameter model for extremely small form factors. MiniCPM-V 4.7 demonstrates strong performance across various benchmarks, including MMBench, MMMU, and OmniDocBench, even when subjected to significant visual compression. The model utilizes techniques like Lookahead Expert Scheduling (LES) and Self-Replay Fine-Tuning to achieve fast inference speeds and retain general capabilities after fine-tuning. AI
IMPACT Enables advanced multimodal AI capabilities on edge devices, potentially accelerating the development of mobile AI applications.
RANK_REASON New open-source multimodal model release from a significant AI lab. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
