llama.cpp b9837 adds --reasoning-preserve flag for chat reasoning traces
New release lets you preserve reasoning in chat templates with Jinja.
The ggml-org/llama.cpp repository released version b9837, featuring a new --reasoning-preserve flag for Jinja-based chat templates. This flag allows models to maintain reasoning traces (e.g., chain-of-thought steps) in conversation history, making it easier for developers to inspect or display the model's internal logic. The feature is implemented in the chat module and comes with a corrected help message, ensuring clarity in CLI usage.
The release spans a wide range of platforms and hardware backends: Apple Silicon (ARM64, with KleidiAI), Intel x64, Ubuntu (CPU/Vulkan/ROCm/OpenVINO/SYCL), Windows (CPU/CUDA 12/13/Vulkan/OpenCL for Adreno), and even s390x mainframes. Android ARM64 and iOS XCFramework builds are also included. This broad support ensures that the reasoning-preserve functionality is accessible across diverse deployment environments, from edge devices to cloud servers.
- Adds --reasoning-preserve flag to Jinja chat templates for retaining model reasoning traces.
- Available across 25+ platform/backend builds including Linux, macOS, Windows, Android, and iOS.
- Supports GPU backends like CUDA 12/13, Vulkan, ROCm 7.2, OpenVINO, and Apple Neural Engine (KleidiAI).
Why It Matters
Enables developers to capture AI reasoning steps in chat apps for transparency, debugging, and advanced interaction flows.