The ggml‑org/llama.cpp project released a new version on 2026-10-21. It replaces a single global scaling factor for mixture‑of‑experts (MoE) down‑projections with a per‑expert‑token factor, using a POOL_1D max reduction. The change also adds support for tensor parallelism by handling POOL_1D in the split‑state logic. The update includes code clean‑ups and a refactored scaling return structure.
Why it matters
Developers running MoE models can now get more accurate scaling on each token, which may improve performance on multi‑GPU setups.
