llama.cpp build b10251 adds Multi-Token Prediction (MTP) support for GLM-4.7-Flash, letting the model generate multiple tokens per forward pass instead of the standard one-at-a-time loop.
The change lands in PR #24868 and ships prebuilt binaries across a wide platform matrix:
- macOS: Apple Silicon (arm64), Intel (x64), iOS XCFramework
- Linux: Ubuntu x64/arm64/s390x CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16
- Windows: x64/arm64 CPU, CUDA 12 (12.4 DLLs), CUDA 13 (13.3 DLLs), Vulkan, OpenVINO, SYCL, HIP
- Mobile: Android arm64, Windows arm64 with OpenCL Adreno
- openEuler: 310p and 910b with ACL Graph
Two platform variants are explicitly DISABLED in this build: macOS Intel with KleidiAI, and the base openEuler target.
MTP is specific to GLM-4.7-Flash in this release. Other models continue using the standard single-token generation path. The llama.app web UI was also updated alongside the core build.
NOTE
Windows ships two separate CUDA builds: one for CUDA 12.4 and one for CUDA 13.3. The DLLs are not interchangeable — pick the package matching your installed toolkit.
Build b10251 is available now on the GitHub releases page, with binaries for every platform listed above.