The llama.cpp project published pre-release b12091 on November 4, 2026. Its main change saves and restores hidden state for MTP, a method used in speculative decoding, when checkpointing. The release notes say it saves only the boundary hidden row; verification rows are rebuilt on supported restore paths.
The release lists builds for macOS, iOS, Linux, Android and Windows, including options for different processors and accelerators. The notes do not describe the change’s performance impact.
Why it matters
For developers using MTP in llama.cpp, the change affects how hidden state is preserved across supported restores; the release notes do not say whether it improves speed or reliability.
