llama.cpp b10217 enables tool calls during DS4 thinking
DeepSeek models can now call tools mid-reasoning on local hardware, changing agent workflows.
llama.cpp, the wildly popular open-source library for running LLMs locally, shipped release b10217 on August 1st. The headline change is a new feature: enabling tool calls within the thinking stage for DS4, likely referring to DeepSeek's latest model family. Previously, reasoning models like DS4 would generate their internal chain-of-thought and only then optionally call tools. With this update, the model can request tool use mid-reasoning, producing more iterative and dynamic agentic behavior. The GitHub commit (ddd4ec1) explicitly states: "chat: enable tool call in thinking for DS4 (#26269)".
This release also showcases llama.cpp's massive ecosystem reach. It ships prebuilt binaries across every major platform: macOS (Apple Silicon with and without KleidiAI), iOS, Linux (x64, arm64, s390x, plus Vulkan, ROCm, OpenVINO, SYCL), Android, and Windows (CPU, CUDA 12/13, Vulkan, OpenCL, HIP). That breadth means developers can deploy DS4 tool-calling on anything from an Android phone to a multi-GPU ROCm server. The project continues to dominate local inference, now with 122k GitHub stars and 21.2k forks. For users, this update closes the gap between cloud-hosted agents and fully local, private reasoning models that can act on their environment.
- llama.cpp b10217 adds tool-call support during the thinking phase for DeepSeek's DS4 model, enabling more iterative agent loops.
- Release includes binaries for macOS, Windows, Linux, Android, and iOS, with accelerators: CUDA 12/13, Vulkan, ROCm 7.2, OpenVINO, SYCL, and OpenCL.
- Project has 122k stars and 21.2k forks, reflecting its status as the standard for local LLM inference.
Why It Matters
Local reasoning agents can now use tools mid-thought, making private, on-device AI far more capable for real-world tasks.