Skip to content
AIBRAINWORX / OpenSourceCoder.comLOCAL-FIRST. BUILT IN THE OPEN.
Linux / Open Source / Developer Tools / Local Inference

llama.cpp b12469 Adds Metal Implicit GEMM Kernel for 2D Convolutions

The ggml-org/llama.cpp pre-release b12469 introduces an implicit GEMM kernel for CONV_2D on Apple's Metal backend alongside cross-platform binary builds and attestations.

Official project or press image for llamacpp-b12469-metal-implicit-gemm-conv2d
Official image from github.com; identification only, no endorsement.
Event / context: 2026-11-17. Source check: October 28, 2026.

The llama.cpp b12469 pre-release from ggml-org adds an implicit GEMM kernel for CONV_2D operations to the project's Metal backend. Published on November 17, 2026, from commit e0a9118, the automated release bundles a single listed code change from pull request #28069 alongside signed build artifacts across desktop, mobile, and server targets.

What changed in the Metal backend

According to the release notes, the lone highlight in tag b12469 is metal : add implicit GEMM kernel for CONV_2D (#28069). In GPU compute backends, implicit general matrix multiplication (GEMM) is a standard technique for mapping two-dimensional convolution operations onto matrix-multiplication hardware paths without materializing a full intermediate im2col buffer in memory.

The release page does not publish throughput numbers, memory deltas, or model-specific benchmarks for the new kernel, so practical performance gains on Apple Silicon hardware remain unverified from the changelog alone. For developers running multimodal or vision-capable models that rely on CONV_2D layers through ggml on macOS or iOS, the update expands native Metal operator coverage directly within the prebuilt distribution.

Prebuilt targets and build scope

Tagged as a GitHub pre-release and signed by GitHub Actions, b12469 includes attestation records and binary archives covering Apple Silicon (arm64), macOS Intel, and iOS XCFramework, as well as Linux, Android, and Windows packages for CPU, Vulkan, CUDA 12 and 13, ROCm 10.0, OpenVINO, SYCL, IBM zDNN, and Qualcomm Snapdragon (CPU, Adreno GPU, and Hexagon NPU) targets.

Two build groups are explicitly marked disabled on the release manifest: the KleidiAI-enabled macOS Apple Silicon arm64 variant and the openEuler target list. Developers evaluating the implicit GEMM convolution path on Apple hardware should treat b12469 as an incremental rolling snapshot and verify behavior against their specific models and Metal runtime environment.

EXPLORE THE SOURCE

GitHub repositories

Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-17.

Follow your curiosity.

Pass the signal.