The llama.cpp project released version b11642 on 2026-10-15. It changes how the server stores prompt‑cache slots. Only slots that can be restored later are saved, so idle slots for rerank, embedding, decision and infill tasks no longer fill the cache. Decision tasks now skip the cache entirely unless the --cache-ram flag is used. This reduces RAM use and avoids unnecessary cache churn.
Why it matters
Developers running llama.cpp on limited hardware will see lower memory usage and fewer cache‑related slowdowns.
