llama.cpp b10486 fixes LFM2 image tiling and Windows bug

llama.cpp, the wildly popular C/C++ implementation for running LLMs locally, has dropped a new release tagged b10486. The project, which boasts 125k stars and 21.9k forks on GitHub, focuses on a targeted bug fix for LFM2 image tiling. The change, contributed via PR #27057 and co-authored by Xuan Son Nguyen from Hugging Face, corrects the threshold used when tiling images for LFM2 models. This is critical for vision-language models that process images in patches; an incorrect threshold can lead to distorted or improperly segmented inputs. The release also refactors testing logic and addresses a Windows-specific issue, making the update more robust across platforms.

What makes this release particularly useful for developers is the extensive prebuilt binary matrix. llama.cpp now offers ready-to-run executables for macOS (both Apple Silicon and Intel), iOS, Linux (x64, arm64, s390x), Android (arm64), and Windows (x64 and arm64). Backend support spans CUDA 12.4 and CUDA 13.3/13.4, Vulkan, ROCm 7.14, OpenVINO, SYCL (FP32/FP16), and OpenCL for Adreno GPUs. This means users can grab a matching binary without compiling from source, regardless of whether they're on a gaming GPU, a data-center accelerator, or a mobile device. For a project that already dominates self-hosted AI, this release keeps llama.cpp a go-to choice for rapidly deploying and testing LLMs and multimodal models locally.