OpenBMB Releases MiniCPM-V 4.7 35B-A3B, First Open-Source MoE Vision Model with Smooth On-Device Inference

OpenBMB has released MiniCPM-V 4.7, introducing the 35B-A3B variant as the first Mixture-of-Experts (MoE) model in the MiniCPM-V series. This release focuses on combining strong visual understanding capabilities with efficient on-device inference, targeting mobile and local computing environments.

On-Device Performance and Efficiency

The 35B-A3B model is designed to run smoothly on-device without compromising performance, achieving a decoding speed of 13+ tokens/s on an iPhone 17 Pro Max. This efficiency is enabled by the new Lookahead Expert Scheduling (LES) technique, combined with fine-grained layer-wise quantization and DSpark speculative sampling integration. In Instruct mode, the model achieves a higher overall score in less inference time compared to Qwen3.6 35B-A3B, advancing the performance-efficiency Pareto frontier for multimodal LLMs.

Architecture and Visual Processing

MiniCPM-V 4.7 builds on the intra-ViT early compression technique from LLaVA-UHD v4, which compresses visual tokens early within the vision encoder to reduce computation in subsequent layers. The model supports mixed 4x/16x visual token compression, allowing developers to balance visual detail against inference costs. It retains the family's capabilities for single-image, multi-image, and video understanding.

Platform Support and Developer Tools

The model offers broad platform support, including iOS, Android, and HarmonyOS, with day-0 support for chip platforms from Rockchip, Houmo, Intel, AMD, and Arm. For developers, MiniCPM-V 4.7 supports inference frameworks such as SGLang, vLLM, llama.cpp, and Ollama, as well as fine-tuning frameworks like SWIFT and LLaMA-Factory. Quantized variants are available in GGUF, BNB, AWQ, and GPTQ formats, with new support for FP8 quantization. Additionally, OpenBMB released Self-Replay Fine-Tuning, a method to extend model capabilities at low cost while minimizing the forgetting of existing skills.