OpenBMB releases MiniCPM V 4.7, open multimodal models built to run on phones and tablets, including a 35B mixture of experts model
🚀 Introducing MiniCPM-V 4.7, our new open-source MLLM series! Advancing the Performance–Efficiency Pareto Frontier. It brings strong multimodal capabilities to iOS and Android devices with extreme efficiency. Two new models: 🔹 MiniCPM-V 4.7 35B-A3B — Bring SOTA performance to local devices with newly introduced LES (Lookahead Expert Scheduling) technique, a new milestone of on-device multimodal AI. 🔹 MiniCPM-V 4.7 1B — Lightweight, efficient multimodal AI. Easier to deploy. ⚡ What's new? SOTA performance: Competitive results across 34 benchmarks spanning multimodal understanding, STEM reasoning, OCR and long-video comprehension. Fast on-device inference: LES predicts and preloads expert weights to reduce inference stalls, boosting 5× speedup on iPad Pro M4 (16GB) with byte-for-byte identical outputs. It enables MiniCPM-V 4.7 35B-A3B to achieve 17+ tokens/s on iPad Pro M4 (16GB) and 13+ tokens/s on iPhone 17 Pro Max, while costing only 3GB memory 🤯 Efficient visual processing: 4× and 16× hybrid visual token compression. Even at 16× compression, 35B-A3B scores 90.4 on OCRBench, 5.2 points ahead of Qwen3.6 35B-A3B. Easier customization: We introduce Self-Replay Fine-Tuning, a simple way to customize new target capabilities at low cost while maintaining general performance. Broad hardware support: Day-0 adaptations for Apple, Intel, AMD, Arm and Rockchip, with many more platforms supported through FlagOS! Download, deploy, and build something we haven't imagined yet. 👀👇 🔗 GitHub: t.co/msjm2QlrvX 🤗 Hugging Face:
Context
Running a capable vision and language model locally on an iPhone or iPad keeps data on the device and avoids cloud costs, which matters for privacy sensitive apps.
OpenBMB is the open source group behind the MiniCPM family of small models, developed with Tsinghua University researchers and the company ModelBest. The new release has two models, a 35B mixture of experts model with about 3B parameters active per token and a 1B model. All performance figures are OpenBMB's own.
Deep dive
On October 12, OpenBMB posted that MiniCPM V 4.7 is its new open source multimodal LLM series, bringing strong multimodal capabilities to iOS and Android devices with extreme efficiency.
According to the post, the larger model uses a new technique called Lookahead Expert Scheduling, which predicts and preloads the expert weights the model will need next, to cut stalls during inference. OpenBMB says this gives a 5 times speedup on an iPad Pro M4 with 16GB of memory, with outputs identical byte for byte, and lets the model reach more than 17 tokens per second on that iPad and more than 13 tokens per second on an iPhone 17 Pro Max while using only 3GB of memory.
The post claims competitive results across 34 benchmarks covering multimodal understanding, STEM reasoning, OCR and long video comprehension. It says that even with 16 times visual token compression, the 35B model scores 90.4 on OCRBench, 5.2 points ahead of Qwen3.6 35B.
OpenBMB also introduces Self Replay Fine Tuning, which it describes as a low cost way to add new capabilities while keeping general performance, and says it has day zero support for Apple, Intel, AMD, Arm and Rockchip hardware. AIReport.NET reported the models are released under the Apache 2.0 license.
What to expect next
Watch for independent benchmarks and on device speed tests, since the 17 and 13 tokens per second figures come from OpenBMB.
Also watch whether app developers ship MiniCPM V 4.7 inside iOS and Android apps, whether Lookahead Expert Scheduling is adopted by other mixture of experts models for phones, and how the release compares with on device models from Apple, Google and Qwen.