llama.cpp b10288 brings back MiniMax M3 support
The 123k-star C++ LLM runner updates tests and ships cross-platform binaries.
llama.cpp, the ubiquitous C/C++ library for running large language models on local hardware, has dropped a new edge release: b10288. Maintained by ggml-org, the project now boasts 123k GitHub stars and 21.4k forks, and its latest build re-enables MiniMax M3 in the test-llama-archs suite via pull request #26633. While this is a small change—focused on test coverage rather than a new feature—it signals that the popular MiniMax M3 model architecture is officially supported and verified in the current codebase, keeping llama.cpp aligned with the latest open-weight models.
For users, b10288 delivers an updated set of prebuilt binaries covering an extensive range of platforms. The release includes macOS Apple Silicon (with and without KleidiAI acceleration), iOS XCFramework, Linux on x64/arm64/s390x, Windows for x64/arm64, Android arm64, and even openEuler with Ascend NPU support. Backends span CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL (FP16/FP32), CUDA 12 and 13, HIP, and OpenCL Adreno for mobile GPUs. This breadth means developers can grab a single archive for their target device and immediately run models like MiniMax M3 locally, without compiling from source. The release follows the project's rolling cadence of regular snapshots, making it easy to stay current with both performance optimizations and emerging model architectures.
- Re-enables MiniMax M3 in test-llama-archs (#26633), confirming official architecture support
- Ships prebuilt binaries for 20+ platform/backend combos, including CUDA 12/13, Vulkan, ROCm, and OpenVINO
- Part of llama.cpp's continuous release stream from the 123k-star ggml-org project
Why It Matters
For developers running local LLMs, llama.cpp's latest build ensures immediate compatibility with new models like MiniMax M3 across every major platform.