TL;DR: The latest llama.cpp release fixes LFM2/LFM2.5 reasoning round-trip formatting issues and token leaks.
Summary: Release b9544 of llama.cpp addresses formatting bugs in common/chat, preventing stray token leaks and resolving round-trip reasoning issues for LFM2 and LFM2.5 models. It ensures reasoning templates correctly gate formats across supported platforms like macOS, Linux, and Windows. The update also includes code cleanup and logging format specifier fixes.
Why it matters: Improperly gated reasoning tokens can break structured output and leak internal thinking processes in reasoning models. Developers using LFM2/LFM2.5 locally should update to prevent formatting errors in chat interfaces.
Source: github.com