The llama.cpp project added support for Soprano-80M and Soprano-1.1-80M in release b11850. These speech-generation models use a Qwen3 backbone and a Vocos audio decoder, which llama.cpp now supports through its multimodal code.
The release notes say the decoder processes the whole utterance at once. In stream mode, llama.cpp therefore sends audio as one chunk after the pipeline is flushed, rather than delivering it piece by piece. The notes document the software support but do not specify model access terms or performance.
Why it matters
Developers using llama.cpp can add these Soprano speech models to their projects, but applications needing audio delivered in small streaming chunks may need a different approach; the release notes do not say how fast generation is.
