OpenBMB's MiniCPM-V 4.6 Runs a Full Vision AI Model on Your Phone
OpenBMB's MiniCPM-V 4.6 packs a 1.3B vision-language model into your phone, outperforming larger rivals with 50% fewer visual encoding FLOPs
Takeaways
- MiniCPM-V 4.6 is a 1.3B vision-language model from OpenBMB that runs fully on-device on iOS, Android, and HarmonyOS with no cloud dependency.
- It scores 13 on the Artificial Analysis Intelligence Index, beating Qwen3.5-0.8B (score 10) using 19x fewer tokens and outperforming Ministral 3B.
- A new intra-ViT early compression technique from LLaVA-UHD v4 cuts visual encoding FLOPs by over 50%, enabling ~1.5x token throughput vs Qwen3.5-0.8B.
- Supports switchable 4x (detail) and 16x (speed) visual token compression modes at inference time.
- The 4-bit quantized version runs on a single RTX 4090 with 6GB VRAM; a GGUF build runs on CPU with ~2GB memory.
- Fully open-source (Apache 2.0) with weights on Hugging Face and edge code on GitHub; supports vLLM, SGLang, llama.cpp, and Ollama.
MiniCPM-V 4.6 is a 1.3-billion-parameter vision-language model from OpenBMB, a Beijing-based AI lab spun out of Tsinghua University, and it does something that still feels a little surreal: it runs fully on-device on an iPhone, an Android phone, or a Huawei HarmonyOS device, with no API call, no cloud round-trip, and no GPU server humming in a data center somewhere. It is available right now, open-source under Apache 2.0, with pre-built apps and fully open-sourced edge deployment code.
The problem it actually solves
Most multimodal models today have one major problem: they are absurdly expensive to run. Even when companies claim a model is "efficient," it often still requires high-end GPUs, massive VRAM, and aggressive cloud scaling. That completely breaks the dream of private, on-device AI. MiniCPM-V 4.6 achieves results on standard benchmarks that previously required models three to five times larger.
The 4.x generation closes the gap on much larger models in specific task categories, particularly OCR, document understanding, and structured data extraction from images. The practical upshot: you can now ship a capable visual AI feature in a mobile app without touching a cloud provider.
The architecture trick that makes it possible
The core innovation is borrowed from a research paper called LLaVA-UHD v4. The key insight is about where you compress visual tokens. Traditional vision-language models run the full Vision Transformer (ViT) on an image first, producing a massive sequence of tokens, and only then compress them down before feeding them to the language model. That means you pay the full quadratic attention cost of the ViT before any reduction happens.
LLaVA-UHD v4 introduces intra-ViT early compression, which reduces tokens in shallow ViT layers and substantially lowers visual-encoding FLOPs while preserving downstream performance. In plain terms: instead of compressing after the expensive part, you compress