MiniCPM-V 4.7: A 35B Multimodal MoE That Runs on Phones With ~3 GB

Big multimodal models usually mean a big GPU or a cloud bill. On October 15, 2026, Beijing-based ModelBest and the OpenBMB community announced MiniCPM-V 4.7, an open-source family built to see images, documents and video directly on phones and tablets. The weights were already posted on GitHub on October 12; today’s press release makes the launch official.

Two sizes: 35B-A3B and 1B

The headline model is MiniCPM-V 4.7 35B-A3B, the series’ first mixture-of-experts (MoE) model. It has 35 billion parameters in total but activates only about 3 billion per token. Most experts stay on the device’s storage and are loaded only when needed, which is how ModelBest gets it running with about 3 GB of memory on iOS and Android. The GitHub README also lists HarmonyOS.

A smaller 1B version builds on MiniCPM-V 4.6 and targets entry-level hardware, with general vision, OCR, chart reasoning and multi-image understanding.

The trick: Lookahead Expert Scheduling

The key engineering piece is Lookahead Expert Scheduling (LES). While one layer computes, LES predicts which experts the next layer will need and loads them in advance. ModelBest says it prepares more than 80% of the required experts ahead of time without extra training, and because the model’s original router still makes the final choice, outputs are identical with LES on or off.

Combined with layer-wise quantization and contiguous expert storage, the company reports:

• about 5x faster average generation on an iPad Pro M4 with 16 GB of memory
• more than 17 tokens per second on that iPad Pro M4
• more than 13 tokens per second on an iPhone 17 Pro Max

What it can do

MiniCPM-V 4.7 covers general visual understanding, STEM problem solving, OCR, document and chart interpretation and long-video understanding. Developers can pick 4x or 16x visual token compression to trade detail for speed; at 16x, ModelBest reports 90.9 on MMBench and 90.4 on OCRBench. A new Self-Replay Fine-Tuning method mixes the model’s own high-quality answers into training data so it can learn new tasks without forgetting old ones, at a cost close to standard fine-tuning.

Hardware support from day one

The release has Day-0 support from AMD, Apple, Arm, Houmo AI, Intel, Rockchip and NVIDIA, plus a long list of Chinese AI and edge chips (Huawei, Hygon, Moore Threads, MetaX, T-Head and others) through the FlagOS stack. Weights, inference code and fine-tuning code are public, with model cards on Hugging Face.

Rocks take

All the speed and benchmark numbers are vendor-reported, so wait for independent tests. Still, this is exactly the direction we want for a local-first smart home: a vision model that can read a utility meter, a parcel label or a camera snapshot without sending anything to the cloud. Rockchip support is the detail to watch — that’s the family of chips inside many single-board computers people already run next to Home Assistant. We’ll try it on our own hardware before recommending it.