
The maintainers of llama.cpp have published pre-release b12457 on GitHub (commit ad5cbfa, released November 17 at 05:20–05:25 UTC, or November 16 in Pacific Time). Tagged as a pre-release and built automatically via github-actions with a verified GitHub signature and linked build attestations, the update centers on a single targeted GPU backend addition: opencl: add A7 Q4_K non-MoE binary kernel (#30948).
What Changed in b12457
According to the official b12457 release notes, pull request #30948 introduces an OpenCL binary kernel tailored for Q4_K quantized, non-Mixture-of-Experts (non-MoE) workloads targeting A7-class hardware. Q4_K is one of llama.cpp's widely used 4-bit K-quant formats for balancing memory footprint and inference quality on constrained devices, and dedicated OpenCL kernels are primarily aimed at improving execution paths on OpenCL-capable GPUs such as Qualcomm Adreno devices where Vulkan or CUDA may not be the preferred runtime.
The release page does not include performance figures, token-per-second comparisons, or independent test results for the new kernel, so real-world speedups or power-efficiency changes on target hardware remain unverified until measured directly on compatible devices.
Cross-Platform Build Matrix and Availability
Alongside the code change, the b12457 tag distributes 40 automated build assets accompanied by GitHub build attestations. Supported packages span macOS (Apple Silicon and Intel) and iOS XCFramework, Linux (Ubuntu x64, arm64, and s390x with CPU, Vulkan, CUDA 12.8, CUDA 13.5, ROCm 10.0, OpenVINO, SYCL, zDNN, and Snapdragon Adreno/Hexagon targets), Android arm64 (CPU and Snapdragon), and Windows x64/arm64 (including OpenCL Adreno, CUDA 12.4/13.5, Vulkan, OpenVINO, SYCL, and ROCm 10.0).
Two build targets appear marked as disabled in the release matrix: the Apple Silicon KleidiAI-enabled macOS build and the openEuler section header (though individual openEuler 310p and 910b entries remain listed in the notes). Because b12457 is an automated rolling pre-release rather than a curated milestone, developers deploying production inference pipelines should evaluate compatibility and verify checksums and attestations before upgrading.
EXPLORE THE SOURCE
Projects in this story
GitHub repositories
- ggml-org/llama.cpp ↗View project details
Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-17.



