llama.cpp b10219 Release Persists Reasoning Content in CLI Chat History
The b10219 release of llama.cpp updates the command-line interface (CLI) to persist reasoning content within the chat history. This allows the CLI to re-inject prior thoughts on subsequent turns when using the `--reasoning-preserve` flag. This update improves the usability of reasoning-focused models like DeepSeek-R1, ensuring they maintain context and logical continuity across multi-turn conversations. Without this, these models would lose their previous chain-of-thought context in subsequent prompts. Previously, `llama-cli` only displayed streamed reasoning content but did not store it in the assistant's message history, rendering the `--reasoning-preserve` flag ineffective. The release also includes pre-built binaries for various platforms, though macOS Apple Silicon with KleidiAI support is currently disabled.
## BACKGROUND
`llama.cpp` is a popular open-source C/C++ implementation for running LLMs locally with high performance. Reasoning models generate a chain-of-thought (reasoning content) to solve complex logical problems before producing their final response.