TL;DR: llama.cpp's latest update fixes an issue where per-request reasoning budget tokens were ignored in chat completions, ensuring caller-supplied values are now honored.
Summary: The llama.cpp b9982 release addresses a bug in its server's chat completions. Previously, the reasoning_budget_tokens field from request bodies was not correctly processed, causing per-request overrides to be discarded. The fix ensures that caller-supplied reasoning_budget_tokens and reasoning_budget_message values are now properly applied.
Why it matters: This update is crucial for developers needing fine-grained control over model 'thinking' time and messages in local LLM deployments. Builders can now reliably customize reasoning behavior per API call, enabling more efficient and tailored AI interactions.
Source: github_releases