llama.cpp performs large language model (LLM) inference in C/C++ and has 131822 stars. Three releases went out during October 20, 2026, from b11753 through b11756, with no separate commit groupings. The three touch different layers. b11753 added \p{M} to the Tekken and GPT-4o pre-tokenizer regexes (the pre-tokenizer is the step that splits text into token units), widening the range of combining character notation handled, and b11755 fixed a missing variant in the Metal MoE (Mixture of Experts, an approach where only some experts take part in the computation for each token) kernel that uses 9 experts per token.
The remaining one concerns server behavior. When Anthropic-spec /v1/messages requests were converted into the internal OpenAI format, id_slot was dropped, and as a result requests that specified a slot holding a KV cache were switched to automatic selection and the cache was recomputed. Other completion endpoints receive id_slot unchanged, so the issue was an asymmetry narrowed to a specific protocol conversion path. Along with the fix, a test was added to check that the specified slot is actually used. The b11753 build also ships binaries for multiple CPU, GPU, and mobile environments.
The grounds are only the change lists in the release notes. The b11755 fix is limited to that kernel handling in the Apple Silicon execution path, and the effect of the b11753 regex expansion on actual tokenization results is not given as comparative figures. No performance or accuracy metrics are included either. So this bundle reads as a tokenizer scope completion and a bug fix for a specific path, and it is too early to judge what ripple effects it has on overall performance.