llama.cpp b10219 fixes reasoning preservation in chat history
llama.cpp's latest release finally persists AI reasoning across turns, fixing issue #26362.
ggml-org pushed llama.cpp b10219, a fresh release of the popular C/C++ inference engine for running LLMs locally. The headline fix addresses a subtle but important CLI issue (#26362): llama-cli collected reasoning_content from the stream for display, but only stored the assistant's final content in message history. As a result, the --reasoning-preserve flag failed to re-inject earlier chain-of-thought tokens on subsequent turns in a conversation.
Now, reasoning is properly persisted in chat history, so multi-turn sessions retain the model's internal thought process for context. This is crucial for anyone building assistants or tools that rely on transparent, step-by-step reasoning. The release also ships a wide range of prebuilt binaries covering macOS (Apple Silicon, Intel), Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Windows (CUDA 12/13, Vulkan, HIP), Android, and iOS, making it easy to update on almost any platform.
- Fixes CLI issue #26362: reasoning_content now stored in chat history
- Enables --reasoning-preserve to re-inject prior thoughts on later turns
- Ships prebuilt binaries for macOS, Linux, Windows, Android, iOS, and more
Why It Matters
For local LLM developers, this enables coherent multi-turn conversations with preserved chain-of-thought, improving agent and assistant reliability.