OpenBMB has published technical details on MiniCPM-V 4.7, a multimodal model that brings a 35-billion-parameter Mixture-of-Experts architecture to the phone. The most striking part is that the model is not shrunk or stripped of experts: the work targets the real bottleneck, the time needed to move expert weights from flash memory to the processor.
How advanced scheduling works
The core technique is called Lookahead Expert Scheduling (LES). The system predicts which experts the next layer may need, preloads their weights from flash storage while current computation is still running, and executes the ready experts first, so the processor does not sit idle waiting for data.
Image: GitHub
This approach is combined with optimized weight layouts, mixed-precision quantization and DSpark speculative decoding, which together reach 20.5 tokens per second on an iPhone 18 Pro Max. The figure is reported by OpenBMB in its technical write-up and has not yet been independently verified.
The same idea of streaming MoE experts from mass storage has been explored elsewhere. Turbo-fieldfare ran Gemma 4 on a Mac with just 2 GB of RAM by streaming experts from the SSD, while Colibrì v1.8.0 treats VRAM, RAM and storage as a single memory hierarchy. LES stands out because it does not change the model's routing: the same experts are used, only in an order that speeds up computation.
A model built for the device
MiniCPM-V 4.7 comes in two versions: a 1-billion-parameter model and a 35B-A3B MoE that activates about 3 billion parameters per token. According to the model card, the larger version runs on iOS and Android with about 3 GB of memory, while the smaller one ranks among the best in its size class for OCR, mathematics and video understanding. It handles single images, multiple images and video.
Image: GitHub
In practice, OpenBMB demonstrates the model reading road signs, parsing complex documents and planning under constraints, all on the phone. Developers can get the weights on Hugging Face, including GGUF format, and the model is compatible with Ollama.
The work shifts the problem: running locally no longer requires a smaller model, but a smarter way of moving data. If the numbers hold outside the lab, a multimodal assistant able to reason about images could work without sending anything to the cloud.
No comments yet. Be the first!