MiniCPM-V 4.6: 1.3B Vision-Language Model Deploys on iOS, Android, and HarmonyOS with 50% Less Visual Compute

Hugging Face·medium signal

OpenBMB released MiniCPM-V 4.6 on May 11, a 1.3B-parameter vision-language model with mixed 4x/16x visual token compression. It outperforms Qwen3.5-0.8B on most VLU tasks and approaches Qwen3.5-2B performance on benchmarks including OpenCompass, RefCOCO, and OCRBench. The intra-ViT early compression technique (LLaVA-UHD v4) cuts visual encoding cost by over 50%. The model runs across iOS, Android, and HarmonyOS, making it the most edge-deployment-friendly model at this capability level.