

The maintainers of llama.cpp have published pre-release build b12488 (commit 10152bc), introducing a targeted CUDA backend change that disables quantized matrix-vector (matvec) fusion across all NVIDIA Pascal-architecture GPUs.
What Changed in b12488
Published via automated GitHub Actions on November 17, 2026 (PST; November 18, 2026 at 00:36 UTC), the release notes highlight a single pull request, #32239: ggml-cuda: disable quantized matvec fusion on all Pascal GPUs. No independent benchmark figures or throughput comparisons are included in the release notes, and the change applies strictly to the ggml-cuda backend path on Pascal-generation hardware.
Why Pascal Support Matters
NVIDIA Pascal cards—such as the GeForce GTX 10-series and Tesla P100/P40 accelerators—remain common in budget local-inference rigs and homelab servers due to their VRAM capacity and secondary-market availability. Because Pascal lacks newer tensor-core instructions found in Volta and later architectures, kernel fusion optimizations in ggml-cuda require careful hardware-specific gating to preserve correctness and stability across older compute capabilities.
Multi-Platform Build Availability
Alongside the CUDA change, release b12488 packages 40 downloadable assets with GitHub build attestations across desktop, mobile, and server environments. Supported targets in the release manifest include macOS (Apple Silicon and Intel) and iOS XCFramework, Linux (x64, arm64, s390x, CUDA 12.8 and 13.5, Vulkan, ROCm 10.0, OpenVINO, SYCL, and Snapdragon NPU/GPU), Android, and Windows (CPU, CUDA 12.4 and 13.5, Vulkan, OpenVINO, SYCL, ROCm 10.0, and OpenCL Adreno), while KleidiAI-enabled macOS builds and openEuler targets are marked disabled in the build matrix.
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 2026-11-18.



