The ggml‑org team released version b11612 of llama.cpp on 14 Oct 2026. The update adds an opt‑in NFC (Normalization Form C) step for Unigram tokenizers, which can improve handling of Unicode text. The change was made with help from Codex and includes a small clean‑up of test files and linter warnings. The release also lists builds for many platforms, from macOS Apple Silicon to Windows with CUDA 13.
Why it matters
Developers running LLaMA models can now get more consistent tokenization of accented or composed characters, reducing unexpected output in multilingual apps.
