b9768

ggml-org/llama.cppb9768Jun 23, 2026by github-actions[bot]same day as b9767

What changed since b9767

medium risk

Adds Granite Speech Plus audio model support with conversion and multi-layer concatenation, plus naming/alignment fixes.

New features

  • Granite Speech Plus audio model is now supported, including a conversion path for its weights.
  • granite_speech handling is extended to concatenate multiple feature layers for the Plus variant.

Improvements

  • Audio feature-layer naming was normalized to plural form throughout conversion and runtime code.
  • mtmd feature_layer usage and naming were aligned consistently across the codebase.
  • Minor logging cleanup using f-strings in the conversion script.

Bug fixes

  • Fixed incorrect singular naming of feature_layers in audio model conversion.
  • Fixed mtmd alignment issues related to feature_layer handling.

Upgrade notes

No CLI or API changes are expected for existing models. If you use Granite Speech audio models, reconvert them with the updated converter so feature_layers naming matches the new runtime expectations; previously converted Plus checkpoints may not load correctly.

Analyzed Nov 1, 2026

Upgrading?

Release notes summary

This release introduces support for the Granite Speech Plus audio model, including conversion capabilities and extended multi-layer concatenation features. Various bug fixes were also applied to naming and alignment.

Key Highlights

  • Add Granite Speech Plus model support
  • Extend granite_speech for plus multi-layer concatenation
  • Fix plural naming for feature_layers
  • Fix mtmd alignment

Release Notes (excerpt)

<details open> model: Granite Speech Plus (#24818) * feat: Add conversion support for Granite Speech Plus Branch: GraniteSpeechPlus AI-usage: full (Bob, OpenCode + Qwen3.6-35b) Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> * feat: Extend granite_speech to support plus multi-layer concatenation Branch: GraniteSpeechPlus AI-usage: draft (Bob, OpenCode + Qwen3.6-35b) Signed-off-by: Gabe Goodhart <ghart@us.ibm.com> * fix(conversion): Fix plural naming for feature_layers for audio…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories