llama.cpp b10270 adds Qwen3-TTS with local voice cloning
llama.cpp's newest release enables local text-to-speech and voice cloning via Qwen3-TTS.
llama.cpp, the GitHub project behind the widely-used local LLM runtime (with over 123k stars and 21.3k forks), just dropped release b10270. This version adds full support for Qwen3-TTS, Alibaba's text-to-speech model, via the project's mtmd multimodal engine. The implementation is substantial: it supports the full pipeline including the text backbone model, speaker encoder, code predictor, and a code-to-waveform stage (renamed from code2wav to GEN_WAV). Developers get new APIs like llama_model_get_tok_embd and mtmd_helper_gen_audio, plus a step_prompt() function for advanced generation control.
The release also includes several notable upgrades. A voice cloning demo is included that preserves KV cache between calls, allowing consistent speaker identity across long generations. Under the hood, it uses ggml_build_forward_select for smarter graph execution and fixes numerous compile and CI issues. As a breaking change, the llama-tts binary has been revamped with updated CLI conventions. Prebuilt binaries are available for macOS (Apple Silicon and Intel), Windows (CPU, CUDA 12/13, Vulkan, OpenCL), Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Android, and even openEuler with Ascend NPUs. The update solidifies llama.cpp's position as the go-to tool for running frontier AI models entirely on local hardware.
- Adds Qwen3-TTS support to llama.cpp's mtmd multimodal pipeline with full text, speaker encoder, and code_predictor handling.
- Introduces new mtmd_gen_audio API and renames code2wav to GEN_WAV, plus a voice cloning demo that preserves KV cache.
- Breaking change for llama-tts binary; prebuilt binaries now cover CUDA, Vulkan, ROCm, OpenVINO, SYCL, Android, and Ascend NPUs.
Why It Matters
Local TTS with voice cloning makes privacy-preserving speech generation accessible on consumer hardware, broadening llama.cpp beyond text inference.