The ggml‑org/llama.cpp project released version b12328 on 2026-11-12. The update changes the backend to prefetch only inputs that were computed by earlier splits, avoiding reads of tensors that will be overwritten later. It also adds a check to skip the MoE (mixture‑of‑experts) cache when the current device cannot run the cached operation, preventing unnecessary data copies.
Why it matters
Developers can expect lower memory traffic and faster inference on hardware that cannot handle MoE workloads.
