llama.cpp b10270: Qwen3-TTS support lands with breaking llama-tts change

llama.cpp b10270 adds Qwen3-TTS support through its mtmd multimodal pipeline — and ships a breaking change to the llama-tts binary that anyone running local text-to-speech needs to handle before upgrading.

The release wires Qwen3-TTS into llama.cpp’s existing multimodal (mtmd) framework rather than building a separate inference path. Key changes:

• Qwen3-TTS support via mtmd — text-to-speech inference runs through the same pipeline used for vision models
• Breaking change to llama-tts binary — existing CLI invocations and scripts need updating
• New mtmd_helper_gen_audio API — programmatic interface for audio generation, including llama_model_get_tok_embd for token embeddings
• Voice cloning pipeline — code2wav conversion to GGUF, speaker encoder loading, and subgraph wiring are functional
• Security fixes included in this build

For production users, the Qwen3-TTS integration means you can run an open-weights TTS model fully locally through the same llama.cpp runtime you’re already using for LLM inference. No separate TTS server, no cloud API call. The trade-off is the breaking change — the llama-tts binary interface changed, so any automation calling it directly will break on upgrade.

Available now in release b10270. If you’re running llama-tts in production, hold off upgrading until you’ve audited the interface changes against your scripts.