llama.cpp build b10259 ships one functional change and a sprawling platform matrix. The feature: models can now reshape tensors during load.
What shipped:
- Tensor reshape at load (#26531) — fixes model architectures that previously failed on shape mismatches
- CUDA 13.3 DLLs for Windows x64, alongside existing CUDA 12.4 builds
- ROCm 7.2 for Ubuntu x64
- Vulkan builds for Ubuntu x64, Ubuntu arm64, Windows x64
- SYCL FP32 and FP16 for Ubuntu x64, plus SYCL for Windows x64
- OpenVINO for both Ubuntu and Windows x64
- OpenCL Adreno for Windows arm64
- KleediAI on macOS arm64: disabled
- openEuler builds: disabled
The tensor reshape change targets a specific failure mode: some GGUF models store weights in shapes that don’t match what the inference engine expects at load time. Before this patch, those models errored out. Now the loader reshapes on the fly — no manual conversion step.
The platform list spans 30+ build artifacts across macOS, iOS, Linux, Android, Windows, and openEuler. CUDA 13.3 support is the newest addition — CUDA 13 shipped weeks ago and llama.cpp already has Windows DLLs ready.
NOTE
KleidiAI (ARM’s ML acceleration library) is disabled in this build for macOS arm64. If you relied on KleediAI acceleration, hold on the previous build until it’s re-enabled.
Grab the build for your platform from the b10259 release page. The tensor reshape fix lands in the model loader — no config change needed, just swap the binary.