b10212
ggml-org/llama.cppb10212Jul 31, 2026by github-actions[bot]same day as b10211
What changed since b10211
b10212 skips loading MTP tensors when the model configuration doesn't use them, reducing memory and load time.
New features
- MTP (Multi-Token Prediction) tensors are now loaded only when the active model configuration actually uses them.
Improvements
- Model loading is faster and uses less memory for architectures that support MTP but don't enable it.
- The conditional MTP loading logic was extended to the remaining MTP-capable model implementations, not just the initial one.
Upgrade notes
This is a drop-in build with no config or API changes; just swap the binaries or rebuild. Users running MTP-capable models without MTP enabled should see lower memory usage and quicker startup. No migration steps are required.
Analyzed Nov 1, 2026
Upgrading?
Release notes summary
This release optimizes model loading by skipping the loading of MTP (Multi-Token Prediction) tensors if they are not used by the current model configuration.
Key Highlights
- Loads MTP tensors only if they are really used
- Skips loading MTP in models that support it but do not use it
Release Notes (excerpt)
<details open> llama : load MTP tensors only if they are really used (#26296) * llama : load MTP tensors only if they are really used * llama : skip loading MTP (if not used) in remaining models that support MTP --------- Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com> </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64) - macOS Apple Silicon (arm64, KleidiAI enabled) DISABLED - macOS Intel (x64) - iOS XCFramework **Linux:**…
Published by ggml-org on GitHub.
Read the full release notesRelated Repositories
:electron: Build cross-platform desktop apps with JavaScript, HTML, and CSS
Godot Engine – Multi-platform 2D and 3D game engine
ClickHouse® is a real-time analytics database management system
Self-hosted game stream host for Moonlight.
DuckDB is a high-performance analytical database system