ModelBest Open Sources MiniCPM-V 4.7 for Use on the Phone Itself
ModelBest has released the next generation of its small multimodal family, shipped as a 35B-A3B version and a 1B version under the MiniCPM name. The larger one is the first mixture of experts model in the line, and the company says it runs on a phone using about 3GB of memory while decoding at more than 13 tokens per second.
The weights, code and documentation are published through the OpenBMB repository on GitHub under the Apache 2.0 licence. The 35B-A3B model activates roughly 3 billion parameters for each token it produces, which is how a model of that total size fits inside a phone's memory budget.
But the claim worth scrutinising is the one about speed. ModelBest attributes the result to an inference technique it calls LES, or lookahead expert scheduling, and describes the release as the first open source mixture of experts model that runs losslessly and smoothly on a device. That last word carries weight, since many small models reach phones only after compression that costs accuracy.
On benchmarks reported by the team, the model scores 91.72 on MMBench, 91.2 on OCRBench and 90.21 on CMMLU. In Instruct mode, ModelBest says it reaches a higher overall score than Qwen3.6 35B-A3B while taking less time to reason. The 1B version improves on its predecessor in basic capabilities and, the company says, leads Qwen3.5 0.8B on both overall score and inference efficiency.
Hands on tests published by QbitAI are more concrete than the benchmark table. The model handled chart reasoning and converted academic papers into Markdown with the phone offline, which is the use case this family is built for, private work on local documents that never leaves the device. The same tests noted that the phone stayed offline for the whole task, which is the point of the exercise.
And the release is aimed past phones. The new model supports iOS, Android and HarmonyOS, and has been adapted for chip platforms from Intel, AMD, Arm, Rockchip and Houmo AI, covering mobile devices and local computers alike.
Alongside the model, ModelBest published a fine tuning recipe it calls Self Replay Fine Tuning, or SRFT. The company says the recipe lets developers add new capabilities at low cost while limiting how much the model forgets what it already knew, a common failure when a small model is taught something new. Developers who want to specialise a model this size usually face a trade off between picking up new skills and keeping old ones.
The release continues a steady cadence for the line, following the previous multimodal version earlier this year and a 2B text model in recent months. Both were pitched at the same audience, developers who want capable models small enough to sit on hardware people already own, without a network connection or a bill for cloud inference.
So the pattern is by now familiar, a Chinese lab shipping a small model whose selling point is that it works offline. What remains untested is whether the efficiency gains hold up outside the vendor's own measurements, since the figures come from ModelBest and its collaborators rather than from an independent evaluation.
ModelBest released the 35B-A3B and 1B versions of its multimodal family under the Apache 2.0 licence
The 35B-A3B version is the first mixture of experts model in the MiniCPM series
ModelBest says it decodes at more than 13 tokens per second in about 3GB of memory on an iPhone 17 Pro Max
Reported scores include 91.72 on MMBench, 91.2 on OCRBench and 90.21 on CMMLU
The model supports iOS, Android and HarmonyOS, plus chips from Intel, AMD, Arm, Rockchip and Houmo AI
Source: Pandaily
