ModelBest (Mianbi Intelligence) and the OpenBMB community formally released the multimodal model MiniCPM-V 4.7 on October 13, shipping two versions at once: 35B-A3B and 1B. The former is the first mixture-of-experts (MoE) model in the MiniCPM-V series, with roughly 35 billion total parameters of which only about 3 billion are activated per inference pass. The latter is aimed at entry-level phones.
The official on-device figures: the 35B-A3B uses about 3GB of memory on an iPhone 17 Pro Max and decodes at more than 13 tokens per second. Code and weights are released under the Apache-2.0 license, according to the footer of the GitHub repository.
Fitting 35B into 3GB by scheduling experts ahead of time
The trouble with MoE models is that the total parameter count is large, yet each step uses only a small slice of it. On a server, everything can sit in GPU memory. On a phone it cannot, so the needed experts must be read in from storage as the computation proceeds, and slow reads cause stutter.
ModelBest's answer is called LES (Lookahead Expert Scheduling). According to the GitHub page, it does three things:
- It predicts routing across layers and preloads the experts the next layer is likely to need.
- For the current layer, it computes first with experts that are already in memory and were also selected by the router.
- It pairs this with weight quantization and a redesigned storage layout to shrink the footprint and speed up reads.
The company says this approach delivers roughly a 5x speedup on real devices. An older technique carries over: the vision side compresses visual tokens with a mix of 4x and 16x ratios, something that already appeared in MiniCPM-V 4.6, released in May.
A third-party developer had earlier gone through the model's configuration file: 40 layers, of which 30 use linear attention and 10 use full attention, with 8 of 256 experts selected per token and a 256K context window. The BF16 weights total about 70.4GB. In other words, the 3GB figure is the runtime footprint after quantization and on-demand loading; running the model unquantized would be too much even for an ordinary PC.
The benchmarks only compare against peers in the same class
ModelBest says the 35B-A3B was evaluated on 34 benchmarks covering general vision, STEM reasoning, OCR and charts, and long video. In Instruct mode, its overall score is higher than that of Qwen3.6-35B-A3B, a model of the same 35B-A3B class, with shorter reasoning time, and on some items it beats Qwen 27B and Gemma-4 31B. The 1B version is positioned against Qwen3.5 0.8B and is said to score better overall and run more efficiently.
All of this is self-reported by the vendor. The results table on the GitHub page is an image, media coverage does not go into the individual scores, and no independent reproduction has surfaced so far.
The release itself had a small subplot. According to a third-party developer's notes, the 35B-A3B weights appeared on Hugging Face on the night of October 6 with no model card, no license and no evaluation data; the documentation was only filled in around October 12.
What stands in the way for developers in China
ModelBest has written a long compatibility list this time. Supported systems cover iOS, Android and HarmonyOS, and the listed chips include Intel, AMD, Arm, Rockchip and Houmo AI. The presence of Rockchip and Houmo AI suggests ModelBest is targeting more than phones: edge boxes, robots and in-vehicle devices in China, which often cannot call a cloud-hosted large model over the network.
Also open-sourced is a fine-tuning recipe called SRFT (Self-Replay Fine-Tuning), which the company describes as a low-cost way to extend capabilities while reducing forgetting of existing ones. For teams that want to customize the model on their own devices, that is more useful than a standalone set of weights.
Several caveats are worth noting:
- The 13 tokens/s figure was measured on the iPhone 17 Pro Max. Some media cite the iPad Pro reaching above 15 tokens/s. The company has not yet published speeds for mid-range Android phones or HarmonyOS devices.
- The 3GB is runtime memory. The model files still take up phone storage, and the exact quantized size should be taken from the officially released files.
- The compatibility list says "supported," and whether the inference frameworks and quantization formats on each platform are ready to use is something developers will have to verify one by one.
Over the past several generations, the MiniCPM-V line has stuck to "small parameter counts, on the phone." Bringing MoE onto the device is, judging from the parameter structure, a different trade between total parameters and phone memory: make the total large while keeping what each step touches at around 3 billion. Two things to watch next: whether LES can hold its speed on devices with less memory, and whether third-party evaluations can reproduce those 34 results.
Sources: OpenBMB GitHub repository for MiniCPM-V (LES design, platform support, license); ITHome; CocoLoop; Phoenix New Media Tech; Hugging Face model page and config file (layer count, expert count, weight size).