OpenBMB's MiniCPM-V 4.7 Runs Full Vision AI on Your Phone at 1B Parameters
OpenBMB's new 1B vision language model pushes the performance-efficiency frontier for on-device multimodal AI, beating Qwen3.5 0.8B while running on phones.
- OpenBMB released MiniCPM-V 4.7 1B, an Apache 2.0 multimodal model for on-device deployment.
- Claims to beat Qwen3.5 0.8B on both overall score and inference efficiency in Instruct mode.
- Uses intra-ViT early token compression from LLaVA-UHD v4 with switchable 4x/16x modes.
- Day-0 support for iOS, Android, HarmonyOS and Rockchip, Houmo, Intel, AMD, Arm chips.
- Ships with FP8, AWQ, GPTQ, GGUF, BNB quantizations plus a DSpark speculative decoding draft.
- Introduces Self-Replay Fine-Tuning (SRFT) to reduce catastrophic forgetting during customization.
OpenBMB has released MiniCPM-V 4.7, a 1B-parameter vision-language model built for phones, tablets, edge hardware, and local servers. It handles images, multiple images, video, OCR, document parsing, and visual grounding within a memory and compute budget suited to mobile devices and NPUs.
OpenBMB reports that the 1B Instruct model outperforms Qwen3.5 0.8B in aggregate benchmark scores and inference efficiency. Those results are vendor-reported, so deployment decisions should account for input resolution, compression settings, quantization, runtime, and target hardware.
Early compression carries the speedup
MiniCPM-V 4.7 applies intra-ViT compression inside the vision encoder, following the approach used by LLaVA-UHD v4. Images normally produce large sequences of visual tokens, and every later transformer layer must process them. Compressing that sequence early reduces work across the remaining encoder layers, especially for high-resolution images and video frames.
Developers can choose between 4x and 16x visual-token compression. The 4x setting retains more visual detail and suits OCR, small text, charts, and fine-grained inspection. The 16x setting lowers token count and latency for captions, broad scene understanding, and video summaries. Production tests should measure both settings because image density and text size can shift the quality threshold.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
