ModelBest and the OpenBMB community have released MiniCPM-V 4.7, a new open multimodal model that comes in two sizes: a 35B mixture-of-experts model called 35B-A3B, and a small 1B model. Both handle images, documents, charts and video, and the weights are already on Hugging Face along with GGUF, FP8, AWQ, GPTQ and int4 versions.
The interesting one is the 35B model. It only activates about 3 billion parameters per token, and most of its experts stay on the device's storage and get loaded when needed. ModelBest says this lets it run on iOS and Android with about 3GB of memory. Its own numbers put it at more than 13 tokens per second on an iPhone 17 Pro Max and more than 17 on a 16GB iPad Pro M4. The trick behind that is something it calls Lookahead Expert Scheduling, which guesses which experts the next layer will need and loads them while the current layer is still computing. The router still makes the final choice, so the output is identical with or without it. Combined with layer-wise quantization and contiguous expert storage, ModelBest says this made average generation about five times faster on the iPad Pro. The 1B model builds on MiniCPM-V 4.6 and is aimed at cheaper, entry-level devices.
Chip support was ready on day one. ModelBest lists AMD, Apple, Arm, Houmo AI, Intel, Rockchip and NVIDIA, and through the FlagOS software stack, chips from CIX Technology, Huawei, Hygon, Iluvatar CoreX, Kunlunxin, MetaX, Moore Threads, Sunrise, T-Head and TsingMicro.
The weights and code are both under Apache-2.0, so commercial use doesn't require a separate license. The project only asks users to fill out an optional registration form.