# llama.cpp update skips LM head for pooled embeddings

> The Oct 13 2026 release adds graph changes that avoid unnecessary LM‑head calculations when only pooled embeddings are needed.

Oossa · 2026-10-13 · https://oossa.com/en/llama-cpp-update-skips-lm-head-for-pooled-embeddings

The llama.cpp library (maintained by ggml‑org) released version b11597 on Oct 13 2026. The update changes the computation graph so that when a context extracts pooled embeddings, the model no longer builds the large LM‑head logits layer. This reduces memory use and speeds up inference for embedding‑only tasks. The change also adjusts related graph nodes to keep the shape consistent.

## The facts

- Release tag b11597 published on Tue Oct 13 2026
- Graph now skips LM‑head logits when pooling type is not NONE

## Why it matters

Developers extracting embeddings can run llama.cpp models faster and on smaller hardware.

## Sources & references

1. [ggml-org/llama.cpp b11597](https://github.com/ggml-org/llama.cpp/releases/tag/b11597) – llama.cpp, 2026-10-13

Last updated: 2026-10-13