llama.cpp b10061 Released with NaN Fix and Multi-Platform Builds

The open-source llama.cpp project, known for enabling efficient local inference of large language models, has released version b10061. This update addresses a stability issue: the initialization of all tensors in the test_dsv4_hc test to avoid NaNs in sentinel tensors. The fix was co-authored by contributor Stanisław Szymczyk and is a minor but important patch for ensuring consistent behavior during model validation.

The release is accompanied by an extensive set of prebuilt binaries covering nearly every major platform. Users can choose from Apple Silicon (both standard and KleidiAI-enabled), Intel Macs, iOS XCFramework, Linux on x64 and ARM with CPU or Vulkan backends, plus specialized builds for ROCm, OpenVINO, and SYCL. Windows users get CPU, ARM, CUDA (v12 and v13), Vulkan, OpenCL for Adreno, and HIP. Android ARM64 and openEuler builds are also provided. The project’s website, llama.app, serves as a central hub for official downloads and documentation.