The ggml‑org team updated llama.cpp to read LFM2‑VL tiling settings from GGUF metadata. This lets the model split images into tiles based on min_tiles, max_tiles, tile_size and max_pixels_tolerance values stored in the model file. The release also bundles pre‑compiled binaries for a wide range of platforms – macOS, Linux, Windows, Android and various hardware accelerators such as CUDA 13, Vulkan and OpenCL.
Why it matters
Developers can now run image‑processing LLMs on more devices without writing custom tiling code.
