OpenBMB Releases MiniCPM-V 4.7 for On-Device Vision AI

OpenBMB has released MiniCPM-V 4.7 1B, an Apache 2.0-licensed multimodal model tailored for deployment on consumer devices such as smartphones, tablets, and local edge servers. Capable of handling tasks like optical character recognition, document parsing, video analysis, and visual grounding, the model ships with immediate day-zero support for iOS, Android, and HarmonyOS alongside hardware architectures from Rockchip, Houmo, Intel, AMD, and Arm. In Instruct mode, OpenBMB reports that MiniCPM-V 4.7 outperforms Qwen3.5 0.8B across aggregate benchmark scores and overall inference efficiency.

To maintain high performance within constrained memory and compute budgets, the architecture integrates intra-ViT early token compression adapted from LLaVA-UHD v4. This technique compresses visual token sequences early within the vision encoder, reducing downstream processing burdens. Developers can dynamically toggle between 4x and 16x compression settings depending on their use case. The 4x configuration preserves high visual fidelity required for small text, charts, and fine-grained inspection, whereas the 16x option significantly cuts token counts and latency to accelerate broad scene description, captions, and video summaries.

For deployment versatility, OpenBMB provides a wide selection of model quantizations, including FP8, AWQ, GPTQ, GGUF, and BNB formats, accompanied by a DSpark speculative decoding draft to speed up generation. Customization is supported through a new method called Self-Replay Fine-Tuning, or SRFT, which mitigates catastrophic forgetting during task-specific adaptation. For machine learning engineers and developers, this release makes high-resolution multimodal processing viable on consumer-grade hardware without relying on cloud infrastructure, provided they calibrate the trade-offs between compression levels and processing runtime for their specific target hardware.