OpenBMB Open-Sources MiniCPM-V 4.7: 35B MoE Multimodal on Device
OpenBMB (ModelBest) has released the open-source on-device multimodal model family MiniCPM-V 4.7, available in two sizes—MiniCPM-V-4.7-35B-A3B and 1B—with all models live on Hugging Face. The 35B-A3B is the family's first MoE model (35B total parameters, 3B active), built for smartphone-class deployment and described in multiple posts as the first open-source on-device MoE multimodal model to hit 13 token/s on an iPhone.
Confirmed
- Models and sizes: MiniCPM-V 4.7 comes in 35B-A3B (MoE) and 1B; the 35B-A3B is an image-text-to-text model
- Key tech: Lookahead Expert Scheduling (LES) expert scheduling, hierarchical fine-grained quantization, and DSpark speculative sampling
- Performance: 13+ tok/s on iPhone 17 Pro Max with only 3GB of memory; 17+ tok/s on iPad Pro M4 (16GB); per @victormustar, a 5x speedup on iPad over the previous version
- Release channels: a Hugging Face collection with GGUF, FP8, GPTQ, AWQ, int4, DSpark, and other quantization formats, plus accompanying tooling (per @XianbaoQIAN, including the relevant deployment toolchain)
Why it matters
- 35B-class multimodal models previously couldn't run in real time on phones; this generation uses a MoE architecture plus system-level optimization (scheduling + quantization + speculative sampling) to push inference into the usable range of 13+ tok/s with only 3GB of memory—meaning flagship phones can run multimodal models locally
- Open-sourcing plus multi-format quantized releases lowers the barrier to on-device deployment; @XianbaoQIAN noted its on-device performance now approaches that of large models, which is especially valuable for privacy-sensitive and offline scenarios
2026-10-13 ~ 2026-10-13 · 7 related posts
Primary sources
- MiniCPM-V 4.7: first open MoE MLLM to run smoothly on-device at 13+ tok/s on iPhone — Skyline34rGt ·
- MiniCPM-V 4.7 ships 35B-A3B and 1B on-device multimodal models punching above frontier baselines — Xianbao_QIAN ·
- OpenBMB ships MiniCPM-V 4.7: a 35B vision model at 13+ tok/s on an iPhone with ~3GB RAM — ondevicedev ·
- [source] MiniCPM-V 4.7: first open MoE MLLM to run smoothly on-device at 13+ tok/s on iPhone — Skyline34rGt · 2026-10-13
- [source] MiniCPM-V 4.7 ships 35B-A3B and 1B on-device multimodal models punching above frontier baselines — Xianbao_QIAN · 2026-10-13
- MiniCPM-V 4.7 collection hits Hugging Face with 35B-A3B and 1B in six quantization formats — Xianbao_QIAN · 2026-10-13
- OpenBMB Releases MiniCPM-V 4.7: On-Device MoE Multimodal With 5x iPad Speedup — victormustar · 2026-10-13
3 near-duplicate retellings: ondevicedev · Med1_Ai · SimplyAnnisa