OpenBMB unveils MiniCPM-V 4.7 open-source multimodal models for on-device AI

OpenBMB, an open-source AI research organization, announced the release of MiniCPM-V 4.7, a new series of multimodal large language models (MLLMs) designed for on-device deployment. The announcement, made via the official OpenBMB account on October 12, 2026, introduces two models: MiniCPM-V 4.7 35B-A3B and MiniCPM-V 4.7 1B. The 35B-A3B model incorporates a newly introduced technique called Lookahead Expert Scheduling (LES), which the organization says predicts and preloads expert weights to reduce inference stalls, achieving a 5× speedup on iPad Pro M4 (16GB) with byte-for-byte identical outputs. OpenBMB claims the model reaches 17+ tokens/s on iPad Pro M4 (16GB) and 13+ tokens/s on iPhone 17 Pro Max, while using only 3GB of memory. The series also features 4× and 16× hybrid visual token compression, with the 35B-A3B model scoring 90.4 on OCRBench at 16× compression, 5.2 points ahead of Qwen3.6 35B-A3B, according to the announcement. OpenBMB says the models deliver competitive results across 34 benchmarks spanning multimodal understanding, STEM reasoning, OCR, and long-video comprehension. The release also introduces Self-Replay Fine-Tuning for low-cost customization and includes day-0 adaptations for Apple, Intel, AMD, Arm, and Rockchip platforms, with additional support through FlagOS. In a separate post, the account atomic_chat_hq reported that MiniCPM-V 4.7 beat Qwen 3.6 35B-3A at a 3D Voxel Japanese Pagoda task while consuming 10% less tokens than Qwen 3.6 35B-3A. The models are available for download via GitHub and Hugging Face, as stated in the announcement.