MiniCPM-V 4.7 is a 35B vision model that OpenBMB says runs on a phone in 3 GB
OpenBMB's first mixture-of-experts MiniCPM-V is Apache 2.0 and downloadable now. The on-device trick is guessing which experts the next layer will need.
- Model
- MiniCPM-V 4.7 35B-A3B
- Maker
- OpenBMB
- Size
- 35B total, about 3B active
- License
- Apache 2.0
- Weights
- Downloadable →
OpenBMB has released MiniCPM-V 4.7, the latest in its line of vision-language models built for phones and other small devices. It reads single images, multiple images, and video. The 35B-A3B version is, in the model card’s words, “the first MoE model in the MiniCPM-V series.”
The claim that will interest builders is on the memory line. OpenBMB says the model decodes at more than 13 tokens per second using 3 GB of memory on an iPhone 17 Pro Max. The full-precision weights are about 70 GB.
How it works
Start with the name. MoE stands for mixture-of-experts. Instead of pushing every token through every weight, each layer holds many small sub-networks, called experts, and a router picks a few of them per token. The model has 35 billion parameters in total, and the “A3B” in the name indicates roughly 3 billion are active for any one token. The card does not give a more exact active count.
That is what makes the phone claim possible at all. If only a fraction of the weights are needed at each step, the rest do not have to be in memory. The catch is that you do not know which fraction until the router decides, and fetching weights from storage in the middle of a token is slow.
OpenBMB’s answer is something it calls Lookahead Expert Scheduling. According to the card, it predicts which experts the next layer is likely to need and starts loading them early. The prediction only steers the prefetching. The actual computation still follows the model’s real routing, and an expert the prediction missed is loaded and run as normal. So a wrong guess costs time, not correctness.
On the vision side, the card credits SigLIP2 for the image encoder and Qwen3.5 and Qwen3.6 for the language backbone. Visual tokens are compressed inside the encoder, and you choose how hard: the default 16x mode merges tokens for speed, and a 4x mode keeps four times as many for finer detail.
The benchmarks
There is no table here, because the source does not give us one. The model card and the project’s GitHub README present every benchmark result as a chart image, with no scores in the text. We do not transcribe numbers off charts.
What OpenBMB says in words is that, in Instruct mode, the 35B-A3B gets a higher overall score in less inference time than Qwen3.6 35B-A3B. That is a vendor claim against a single comparison model, the same family its backbone is credited to. We could not find independent results yet.
The numbers that are in the text are these, all from OpenBMB:
| Claim | Figure | Where |
|---|---|---|
| Decoding speed on iPhone 17 Pro Max | 13+ tokens/s | Model card |
| Memory use on that phone | 3 GB | Model card |
| Speedup from the scheduling work on real devices | 5x | GitHub README |
| Size of the BF16 weights | 70.6 GB | Hugging Face file listing |
Loose ends
- Which build runs on the phone. The 3 GB figure cannot be the 70 GB BF16 weights held in memory. The card mentions weight quantization and an on-device app in a separate repository, but does not say in text which quantized build produced the phone number.
- Context window. Not stated on the card.
- The 5x speedup. It appears in the GitHub README. On the model card the before-and-after comparison is an animation with no figure attached.
- Dates. The GitHub README dates the release to October 12. The weight files on Hugging Face were committed several days before that.
Try it yourself
The weights are on Hugging Face under Apache 2.0, as 16 safetensors files with no access gate that we could see. The card lists Transformers 5.18.0 or later, vLLM, SGLang, Ollama, llama.cpp, and LM Studio as ways to run it, and links FP8, int4, AWQ, GPTQ, and GGUF versions for smaller machines. There is also a much smaller 1B sibling.
We have not run MiniCPM-V 4.7 ourselves; this is based on the model card, the Hugging Face file listing, and the project’s GitHub README.