llama.cpp b10217 ships one functional change: tool calls now execute inside DeepSeek thinking blocks. PR #26269 enables the DS4 chat template to parse tool calls during reasoning, not just in the final response.
What’s in the build:
- Headline change: Tool call support in thinking blocks for DS4 template (#26269)
- CUDA 13.3 DLLs now available alongside CUDA 12.4 for Windows x64
- ROCm 7.2 Linux build included
- KleidiAI macOS build disabled this release
- openEuler builds disabled this release
The practical impact: if you’re running a DeepSeek reasoning model locally and want it to call functions mid-reasoning — checking a value, looking something up before continuing its chain of thought — the server now handles that without dropping the call. Previously, tool calls embedded in thinking output were ignored or caused parse errors.
NOTE
“DS4” refers to the DeepSeek chat template version in llama.cpp. The change applies to any model using that template, including DeepSeek-R1 and its distills.
Platform coverage remains broad: Vulkan builds for Ubuntu x64/arm64 and Windows x64, SYCL FP32/FP16 for Intel GPUs, OpenVINO for both Linux and Windows, and OpenCL Adreno for Windows arm64.
Build b10217 is available now across 20+ platform targets. The CUDA 13.3 Windows build is the one to grab if you’re on recent NVIDIA drivers.