llama.cpp Release b10270 Adds Support for Qwen3-TTS Voice Cloning
The llama.cpp project has released version b10270, introducing native support for Alibaba's Qwen3-TTS model. This update enables local text-to-speech (TTS) and voice cloning capabilities, though it introduces breaking changes to the llama-tts binary. By integrating Qwen3-TTS, llama.cpp expands its utility beyond text generation into high-quality, multilingual audio synthesis and voice cloning on consumer hardware. This lowers the barrier for developers looking to deploy private, local, and interactive voice agents. The implementation involves converting multiple sub-components of the Qwen3-TTS pipeline—including the text model, encoder, speaker encoder, and code2wav generator—into the GGUF format. It also introduces a revamped llama-tts binary and a new mtmd_helper_gen_audio API.
## BACKGROUND
llama.cpp is a popular open-source inference engine designed to run large language models locally with minimal setup and high performance, often utilizing quantized GGUF files. Qwen3-TTS is an open-source text-to-speech model family developed by Alibaba Cloud, supporting expressive speech generation, voice cloning, and multilingual synthesis across ten languages.