MiniCPM-V 4.7
ModelBest released MiniCPM-V 4.7, its first mixture-of-experts multimodal model, in 35B-A3B and 1B sizes, with the MoE version running in about 3GB on an iPhone 17 Pro Max.
The 35B-A3B version is the first MoE model in the MiniCPM-V series and activates about 3 billion parameters per token. A new inference technique called lookahead expert scheduling (LES) lets it decode at over 13 tokens per second in about 3GB of memory on an iPhone 17 Pro Max, the company says. The 1B version improves on MiniCPM-V 4.6, and ModelBest also open-sourced Self-Replay Fine-Tuning (SRFT), a recipe for adding capabilities while limiting forgetting.
- Date
- Wednesday 14 October 2026
- Lab
- ModelBest
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| MMBench (35B-A3B) | 91.72 | company |
| OCRBench (35B-A3B) | 91.2 | company |
| CMMLU (35B-A3B) | 90.21 | company |
| Decode speed on iPhone 17 Pro Max (35B-A3B) | >13 tokens/s in about 3GB of memory | company |
Weights are under Apache-2.0 via the OpenBMB GitHub repository. Supports iOS, Android and HarmonyOS. Comparisons with Qwen3.6 35B-A3B and Qwen3.5 0.8B are company claims.
Sources
This record was checked against its sources on 14 October 2026. How we check