b8887
The latest update enables 2x faster performance on Apple Silicon and adds support for Intel's OpenVINO toolkit.
The open-source community behind llama.cpp has rolled out a significant new release, version b8887, expanding the tool's hardware compatibility and performance profile. The update introduces experimental KleidiAI acceleration for Apple Silicon Macs, which early benchmarks suggest can double inference speeds for compatible models. It also adds official support for Intel's OpenVINO toolkit, giving developers optimized performance on Intel CPUs. This release continues llama.cpp's mission of making large language models accessible and efficient across the widest possible range of consumer and server hardware.
The b8887 build significantly broadens platform support, adding pre-built binaries for Android arm64, Windows with CUDA 12.4 and 13.1 DLLs, and multiple Linux backends including Vulkan, ROCm 7.2, and OpenVINO. For enterprise environments, it now includes builds for openEuler with Huawei Ascend 310P and 910B AI accelerator support. The release also fixes a rope type handling issue (#22242) and maintains the project's signature small footprint and efficiency. These updates make llama.cpp an even more versatile tool for developers deploying LLMs in production across cloud, edge, and mobile environments.
- Adds KleidiAI backend for Apple Silicon, promising up to 2x faster inference speeds
- Introduces official OpenVINO support for optimized performance on Intel CPUs and integrated graphics
- Expands platform coverage to 28 pre-built binaries including Android, Windows CUDA 12/13, and openEuler with Ascend AI chips
Why It Matters
Democratizes efficient LLM deployment by optimizing performance across Apple, Intel, NVIDIA, AMD, and mobile hardware.