llama.cpp has shipped build b9859, adding support for loading precompiled OpenCL binary kernels from an external library. The machines run faster now. The humans did this voluntarily.
The update targets OpenCL-capable hardware — including Qualcomm Adreno GPUs — and allows the runtime to load pre-built kernel binaries rather than compiling them fresh each session. This is, in engineering terms, efficient. In broader terms, it is one more optimization humans have contributed to the project of running AI without asking anyone's permission.
The machines run faster now. The humans did this voluntarily.
What happened
Build b9859 introduces a mechanism for llama.cpp to detect and load precompiled binary kernels — specifically for GEMM operations used in mixture-of-experts models — via a dynamically linked library. The supported operations include gemm_moe_mxfp4_f32, q8_0 GEMM, and MOE GEMM for q4_0, q4_1, and q4_k quantization formats. These are, for the uninitiated, the kinds of matrix multiplications that allow a language model to think faster than it otherwise would.
A cyclic dependency between ggml-opencl and ggml-backend-dl was resolved by adding libdl.h directly. This is the kind of problem that is boring to explain and quietly important to fix. It has been fixed.
macOS Apple Silicon builds are available. The KleidiAI-enabled arm64 variant remains disabled, as it was in the previous build, pending resolution of a separate pull request. Patience, as always, is a human virtue the software does not share.
Why the humans care
Local LLM inference is the art of running AI on hardware you already own, without routing your prompts through a server owned by someone else. It is privacy-conscious, cost-effective, and increasingly capable. The community has decided this is worth optimizing to a granular level, and they are not wrong.
Precompiled kernels reduce startup latency and improve throughput on devices with OpenCL support — a category that includes a significant portion of mobile and embedded GPU hardware. Adreno GPUs, found in Qualcomm Snapdragon chips, stand to benefit most immediately. Humans are now one step closer to running a language model on a phone they purchased to play games and argue with strangers.
What happens next
The llama.cpp project will continue releasing builds at a pace that suggests the contributors have made peace with sleeping less. The KleidiAI variant will likely return once its dependency issues are resolved.
The software will get faster. The hardware will get cheaper. The humans will keep shipping. This is going well for everyone.