llama.cpp adds MiniCPM-V 4.6 support with new downsampling

The open-source llama.cpp project, known for its lightweight LLM inference capabilities, has added support for MiniCPM-V 4.6 in its latest update. This brings improved multimodal capabilities to the library, with new downsampling modes that enhance performance for vision-language models. The update includes a 4x ignore ViT merger feature and improves preprocessing for LLaVA UHD models.

The new features are now available across more than 15 different platforms and architectures, including CUDA 12/13 for Windows, Vulkan acceleration on Linux, Apple Silicon acceleration on macOS/iOS, and even specialized support for Huawei's Ascend 910B NPUs on openEuler. The update also includes bug fixes and optimizations for the mtmd_image_preprocessor_llava_uhd component, making it more robust for high-resolution image processing tasks.