On November 10, ModelBest released MiniCPM-V 4.7, a 35-billion-parameter (about 3 billion activated) multimodal model that runs lossless and smoothly on phones, using a mixture-of-experts architecture with about 3 GB memory and over 20 tokens per second generation; a 1-billion version reaches budget and even some entry-level phones.
Key to fitting and speeding the model is storing most expert weights in phone flash storage and loading on demand, with look-ahead expert scheduling hiding storage I/O waits, plus layered quantization and speculative decoding. The breakthrough shows intelligence migrating toward personal devices for privacy, offline use and low cost.
ModelBest has open-sourced the inference engine, app code and packages so phone, chip and app developers can keep optimizing for real scenarios.
Source: Science and Technology Daily / ModelBest
