MiniCPM-V 4.7: Open 35B MoE Multimodal Model Runs on Phones with 3 GB RAM
OpenBMB and ModelBest have released MiniCPM-V 4.7, an open-weight multimodal model family under the Apache 2.0 license designed to run directly on consumer devices. The flagship model, MiniCPM-V 4.7-35B-A3B, is a 35-billion-parameter Mixture-of-Experts (MoE) architecture where only 3 billion parameters are active per token. In edge tests on an iPhone 17 Pro Max, the model runs completely offline, decoding at over 13 tokens per second while fitting into roughly 3 GB of memory.
How the 35B-A3B architecture fits on mobile hardware
Fitting a 35-billion-parameter model into phone hardware usually runs into memory bandwidth walls. MiniCPM-V 4.7 addresses this with Lookahead Expert Scheduling (LES), an algorithm that predicts which expert weights will be required in upcoming steps and streams them ahead of time. Paired with fine-grained layer-wise quantization and speculative decoding, it keeps the memory footprint near 3 GB without sacrificing interactive responsiveness.
Alongside the flagship, the developers introduced MiniCPM-V 4.7-1B, a 1.3-billion-parameter dense model built for constrained environments. It delivers leading marks in optical character recognition, mathematical reasoning, and video understanding within its size category.
Early visual token compression
Another architectural highlight is intra-ViT early visual token compression, derived from the LLaVA-UHD v4 lineage. The model provides adaptive 4x and 16x visual compression directly inside the visual encoder.
At 16x compression, the computational load in later attention layers drops substantially while benchmark metrics hold firm: MMBench drops only from 91.72 to 90.87, and OCRBench shifts from 91.20 to 90.40. In document parsing, the model posted a score of 94.24 on OmniDocBench v1.6, outpacing other compact models by more than 12 points.
Ecosystem and platform support
Day-0 software support includes integrations with llama.cpp, Ollama, vLLM, and SGLang. Pre-built mobile packages are ready for iOS, Android, and HarmonyOS, with hardware acceleration for Intel, AMD, Arm, Rockchip, and Houmo chipsets. For fine-tuning, the team introduced Self-Replay Fine-Tuning (SRFT) to protect pre-trained foundational knowledge when tailoring the model to niche workflows.
What our platform aggregates show
The demand for nimble, responsive models is evident in our production data as well. Over the past 30 days (14.09–14.10.2026), 14,515 generation tasks completed across the platform: 10,879 images (10,145 successful, 93.3%), 3,194 videos (2,243 successful, 70.2%), and 442 music tracks (437 successful, 98.9%). These represent anonymized service aggregates across all users with no personal identifiers.
Among image generation models, fast lightweight options see strong user engagement. For example, the compact Nano Banana 2 handled 2,069 tasks with a 98.0% success rate (2,027 completed runs). You can explore quick multimodal models in our chat.
MiniCPM-V 4.7 illustrates how quickly the boundary between massive server clusters and pocket-sized devices is fading. Offline execution with a 3 GB memory footprint brings private, on-device intelligence closer to daily computing reality.