ModelBest Open-Sources MiniCPM-V 4.7, First MoE Multimodal Model Running in 3GB on a Phone

ModelBest released MiniCPM-V 4.7 in 35B-A3B and 1B sizes under the Apache-2.0 license. The 35B-A3B is the first mixture-of-experts model in the MiniCPM-V line: it activates about 3 billion parameters per token and runs on an iPhone 17 Pro Max using roughly 3GB of memory at over 13 tokens per second via its LES technique. Reported scores include 91.72 on MMBench, 91.2 on OCRBench and 90.21 on CMMLU.

• 35B-A3B activates ~3B parameters per token and needs ~3GB of memory
• Decoding on iPhone 17 Pro Max exceeds 13 tokens per second
• Benchmarks: MMBench 91.72, OCRBench 91.2, CMMLU 90.21
• Supports iOS, Android, HarmonyOS and Intel, AMD, Arm, Rockchip, Houmo AI chips