b10270

ggml-org/llama.cppb10270Aug 4, 2026by github-actions[bot]same day as b10269

What changed since b10269

high risk

b10270 adds Qwen3-TTS speech synthesis to the multimodal stack and reworks the llama-tts binary, a breaking CLI change.

Breaking changes

  • The llama-tts binary interface was revamped, so existing scripts invoking it with the old arguments or output conventions will need updating.
  • Internal graph naming changed from CODE2WAV to GEN_WAV, which may affect tooling or docs that referenced the old identifier.

New features

  • Qwen3-TTS text-to-speech models are now supported through the mtmd multimodal framework, including speaker encoder and code-predictor components.
  • New audio generation helpers in the mtmd API (mtmd_helper_gen_audio) plus a model capability check (mtmd_helper_model_can_chat).
  • New llama_model_get_tok_embd accessor exposes token embeddings to callers.
  • TTS generation now supports sequence IDs and a step_prompt() flow for incremental prompting.
  • Language selection for TTS uses ISO 639-1 codes, with accompanying guide and developer documentation.

Improvements

  • The code2wav stage preserves its KV cache across calls, reducing recomputation during audio decoding.
  • A fused snake activation op and use of ggml_build_forward_select streamline the audio graph.
  • The llama-tts binary was cleaned up and now reports more detailed timing information.
  • Several security fixes were folded into the TTS/multimodal code path.
  • Documentation and dev docs were expanded to cover the new TTS workflow.

Bug fixes

  • Fixed clamping of the cold prefix during audio generation.
  • Resolved compile and lint failures introduced while wiring up the TTS pipeline.

Upgrade notes

If you script the llama-tts binary, review its new argument and output format before upgrading, since the interface changed. Qwen3-TTS requires converting the text model, encoder, speaker encoder, and code predictor to GGUF; follow the new guide. The previous release's dflash wo_a load fix is retained, so no action is needed there.

Analyzed Oct 31, 2026

Upgrading?

Release notes summary

Adds support for Qwen3-TTS (Text-to-Speech) in the multi-modal framework.

Key Highlights

  • Qwen3-TTS support
  • Revamped llama-tts binary
  • Multi-modal generation API updates

Release Notes (excerpt)

<details open> mtmd: support Qwen3-TTS (note: breaking change to llama-tts binary) (#26254) * convert text model * main model load ok * convert encoder ok * speaker encoder loading ok * speaker enc graph * adapt vocab for backbone (with some tricks) * add suppress_tokens * poc new mtmd gen api * convert code_predictor to gguf * load gen_code model ok * add clip_encode * wire up * code gen cgraph init version Co-authored-by: Pascal <admin@serveurperso.com>…

Published by ggml-org on GitHub.

Read the full release notes

Related Repositories