llama.cpp Release b11447 Adds Support for Perplexity's pplx-decider Models
The open-source LLM inference engine llama.cpp has released build b11447, introducing support for Perplexity AI's pplx-decider model architecture. The update also refreshes pre-compiled binary packages across Linux, Windows, macOS, Android, and Snapdragon platforms. Support for pplx-decider enables developers to run Perplexity's decision-making models locally using llama.cpp's high-performance CPU and GPU backends. This highlights llama.cpp's ongoing commitment to quickly integrating emerging fine-tuned architectures into its ecosystem. The change is tracked under pull request #30044 and is included in pre-built release binaries for CUDA 12/13, Vulkan, ROCm, OpenVINO, and SYCL. Note that the macOS Apple Silicon build with Arm KleidiAI acceleration enabled remains temporarily disabled in this release tag.
## BACKGROUND
llama.cpp is a C/C++ LLM inference framework built for low-latency, cross-platform execution of quantized language models. Perplexity AI's pplx-decider models (such as pplx-decider-v1-27b) are decision-making fine-tunes built on architectures like Qwen to evaluate and select optimal reasoning or query routing strategies.