The ggml‑org/llama.cpp project released version b11776 on 2026-10-21. The update adds support for returning log probabilities (logprobs) while running speculative decoding, a technique that speeds up generation by guessing tokens ahead of time. A regression test was also added to ensure the new logprob values match a run without speculation. The fix resolves a bug where the master branch returned empty logprob data.
Why it matters
Developers can now see confidence scores for each token when using fast speculative decoding, helping them debug and improve generation quality.
