The latest llama.cpp release, build b10217, enables tool calling during the thinking phase for DeepSeek (DS4) models in chat mode. This enhancement allows the model to seamlessly invoke external tools while it is still reasoning, producing more coherent and effective multi-step interactions. For self-hosted users running DeepSeek, this means smarter assistant behavior and improved reliability when the model needs to fetch data or perform actions as part of its response generation.
Pre-built binaries are available for all major platforms, including macOS, Windows, Linux, Android, and iOS, covering CPU and various GPU backends such as CUDA, Vulkan, ROCm, and OpenVINO. Download from the official release page to try out the new capability.
