b10444

ggml-org/llama.cppb10444Aug 15, 2026by github-actions[bot]same day as b10443

What changed since b10443

medium risk

b10444 adds --models-dir loading for MTP assistant models and drops eagle3 draft-model support.

Breaking changes

  • Support for the eagle3 draft model type has been removed, so configurations or presets relying on eagle3 will no longer work.

New features

  • MTP (Multi-Token Prediction) assistant models can now be loaded through the --models-dir flag.
  • Preset handling now detects MTP models using a strict prefix check, reducing mismatched model selection.
  • Draft-type handling was generalized to cover additional speculative decoding variants beyond the previously supported ones.

Improvements

  • Preset logic was refactored to reuse newer common infrastructure, simplifying model resolution.
  • General code cleanup and refactoring in the common/preset code paths.

Upgrade notes

If your setup uses eagle3 draft models, migrate to a supported draft type (e.g. MTP) before upgrading, since eagle3 is no longer recognized. Users loading MTP assistant models can now point --models-dir at them, but verify preset prefix matching resolves the intended model. No config migration is required otherwise.

Analyzed Oct 24, 2026

Upgrading?

Release notes summary

This release adds support for loading MTP (Multi-Threaded Processing) assistant models using the --models-dir flag. It includes updates to presets for strict MTP model prefix checking and removes support for the eagle3 model.

Key Highlights

  • Added --models-dir support for MTP assistant models
  • Updated presets to check for MTP models with strict prefix
  • Removed support for eagle3 model
  • General code cleanup and refactoring

Release Notes (excerpt)

<details open> common: support --models-dir loading MTP assistant models (#24431) * common: support --models-dir loading MTP assistant models * common: preset: check for MTP models with strict prefix * common: preset: Take advantage of PR #27005 * handle other draft types * drop eagle3 * clean up --------- Co-authored-by: Xuan Son Nguyen <son@huggingface.co> </details> **Website:** - <https://llama.app> **macOS/iOS:** - macOS Apple Silicon (arm64)…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories