llama.cpp b10093 adds DeepSeek-v4 support with reasoning flag

llama.cpp, the popular open-source C++ library for local LLM inference, has released version b10093 with key support for DeepSeek's v4 model family. The release fixes the DeepSeek4 template to explicitly follow reference behavior, ensuring correct formatting when using DeepSeek v4 models locally. A new `drop_reasoning` flag has been added, giving users control over reasoning token generation—critical for tasks like chain-of-thought prompting.

Additional improvements include hooking the DeepSeek3.2 parser to work seamlessly with v4, and a fix for tool result reordering to maintain response consistency. Build assets are available across macOS, iOS, Linux, Windows, and Android with various backends (CPU, Vulkan, CUDA, ROCm, etc.). This update maintains llama.cpp's position as a leading solution for running state-of-the-art models on consumer hardware.