OpenBMB · 2026-05-11 · major
OpenBMB MiniCPM-V 4.6 — 1.3B Vision-Language Model Runs on iOS, Android, and HarmonyOS, Beating Qwen 3.5-0.8B at 19× Fewer Tokens
Open-source 1.3B multimodal model combining SigLIP2-400M vision and Qwen3.5-0.8B language. Scores 13 on Artificial Analysis Intelligence Index, surpassing Qwen 3.5-0.8B (10) with ~19x fewer output tokens. Ships with edge code for iOS, Android, HarmonyOS.
Pocket-sized vision-language model — 1.3B params, runs on phones, beats larger models per token spent.
Key specs
| License | Apache-2.0 |
|---|---|
| Active params | 1.3B |
| Context window | 262K tokens |
| GitHub stars | 24.6K |
| Intelligence index | 13 |
| Visual flops cut | >50% |
What is it?
MiniCPM-V 4.6 is a 1.3B-parameter open-weights multimodal model from OpenBMB. It takes text, single or multiple images, and video as input and produces text. The repo ships with full edge-adaptation code for iOS, Android, and HarmonyOS, plus support for vLLM, SGLang, llama.cpp, and Ollama.
How does it work?
Vision is handled by SigLIP2-400M and language by a Qwen3.5-0.8B backbone, with mixed 4x/16x visual-token compression on top of LLaVA-UHD v4's intra-ViT early compression cutting visual-encoding FLOPs by more than half. Two thinking modes — fast non-thinking and a longer reasoning mode — are toggled at inference. Output supports GGUF, AWQ, GPTQ, and BNB quantizations for phone deployment.
Why does it matter?
Scores 13 on the Artificial Analysis Intelligence Index — higher than Qwen 3.5-0.8B (10) while burning ~19x fewer output tokens to do it, and ~43x fewer than Qwen 3.5-0.8B-Thinking (11). For mobile and edge ML teams that wanted MLLM capability without a 7B model and a cloud bill, the math finally works.
Who is it for?
mobile devs and edge ML teams
Try it
ollama run openbmb/minicpm-v4.6