Llama.cpp adds support for GLM-4.7-Flash with MTP
Llama.cpp expands GPU/CPU support for GLM-4.7-Flash across 20+ platforms...
The Llama.cpp project (github.com/ggerganov/llama.cpp) has released build **b10251**, introducing support for **GLM-4.7-Flash** with Multi-Token Prediction (MTP) capabilities. This model, developed by **Zhipu AI**, now benefits from optimized inference across an unprecedented range of hardware platforms—covering **20+ CPU/GPU configurations** including CUDA 12/13, Vulkan, ROCm 7.2, OpenVINO, SYCL, and HIP accelerators.
The update prioritizes **cross-platform deployment**, with official builds for macOS (Apple Silicon/Intel), Linux (Ubuntu x64/arm64/s390x), Android (arm64), Windows (x64/arm64), and even niche distributions like openEuler. Notably, Vulkan and OpenVINO backends enable efficient inference on integrated GPUs and NPUs, while CUDA support extends to both 12.4 and 13.x DLLs. Developers can also leverage ROCm for AMD GPUs or SYCL for Intel integrated graphics, making this one of the most hardware-agnostic releases in Llama.cpp’s history.
- Llama.cpp b10251 adds native support for GLM-4.7-Flash with MTP (Multi-Token Prediction) across 20+ CPU/GPU platforms
- Includes optimized builds for CUDA 12/13, Vulkan, ROCm 7.2, OpenVINO, SYCL, and HIP accelerators
- Supports macOS, Linux, Android, Windows, and openEuler with hardware-specific optimizations
Why It Matters
Broadens access to GLM-4.7-Flash for edge/on-prem deployments, reducing vendor lock-in for AI inference.