Reminder: try probabilistic MTP if you missed it. Decode +14% on prose
retained text
[https://github.com/ggml-org/llama.cpp/pull/27694](https://github.com/ggml-org/llama.cpp/pull/27694) Now merged. Update your llama if you haven't done so yet. Optimal draft-n-max / draft-p-min seem to be in line with greedy sampling. Main gain seems to be on **prose generation.** Tests above ran with thinking off, ngram-mod off.
document fetched 2026-10-12 00:37 · canonical https://reddit.com/r/LocalLLaMA/comments/1x33p4o/reminder_try_probabilistic_mtp_if_you_missed_it