OpenBMB releases MiniCPM-V 4.7, an open vision model that runs on phones

China-based OpenBMB released the MiniCPM-V 4.7 models on Hugging Face on October 6 and 7, 2026. The MiniCPM-V series consists of vision language models that understand a single image, several images or a video and answer questions about them in text. The new generation comes in two sizes: 35B-A3B, which has 35 billion parameters in total but runs about 3 billion at each step, and the 1-billion-parameter 1B. Both are licensed under Apache 2.0, so they can be downloaded and run on your own server or device, including for commercial use.

How a large model fits on a phone

35B-A3B is the first MoE (mixture of experts) model in the series. In this design the model uses only some of its experts for each word, yet the files for all experts still have to live on the device. When a phone's memory cannot hold them all at once, experts are reloaded from storage again and again and the model slows down. According to the model card, OpenBMB addresses this with three techniques: Lookahead Expert Scheduling (LES), which predicts and preloads the experts the next layer will need; layer-wise quantization that compresses the expert files; and DSpark, which speeds up generation through speculative sampling. The predictions only guide preloading; the experts the model selects stay the same.

According to screen recordings shared by the company, 35B-A3B generates more than 13 tokens per second on an iPhone 17 Pro Max with about 3 GB of memory, and there are recordings running on an iPad M4 as well. OpenBMB says that in Instruct mode 35B-A3B reaches a higher overall score than Qwen3.6 35B-A3B in less time, and that the 1B version beats Qwen3.5 0.8B on both score and speed.

Supported platforms and tools

The models run on iOS, Android and HarmonyOS, with day-one support for Rockchip, Houmo, Intel, AMD and Arm chips. On computers and servers they work with SGLang, vLLM, llama.cpp and Ollama, and compressed versions such as GGUF, AWQ, GPTQ and FP8 are ready. Images can be compressed 4x or 16x to trade visual detail against speed. OpenBMB also released Self-Replay Fine-Tuning (SRFT), a method meant to reduce the loss of existing skills when you train the model on your own data.

The comparisons were run by OpenBMB, and the model card gives no separate results by language. Before using it on your own documents, screenshots or videos, test it with your own samples. For a fuller picture, follow evaluations from independent researchers as well.

Sources: Hugging Face: MiniCPM-V 4.7 35B-A3B model card, Hugging Face: MiniCPM-V 4.7 1B model card, GitHub: MiniCPM-V.