TL;DR
MiniCPM-V 4.6 is an ultra-lightweight multimodal LLM with 1.3B parameters, efficiently understanding images and videos on smartphones while surpassing larger models in performance.
Key features
Ultra-efficient visual encoding: intra-ViT early compression technique (LLaVA-UHD v4) reduces visual encoding cost by over 50%.
Mixed compression rates: flexible selection of 4x/16x visual token compression rates.
Mobile deployment: runs on iOS, Android, HarmonyOS, etc., with edge adaptation code open-sourced.
High performance: with 1.3B parameters, outperforms larger models like Gemma4-E2B-it, and achieves 1.5x higher token throughput than Qwen3.5-0.8B.
When to use it
When real-time image/video understanding is needed on smartphones or edge devices.
When developing mobile apps that require local multimodal inference without cloud dependency.
When leveraging high-performance vision-language models with limited computing resources.