Open-source full-duplex speech model caramel-1: 413M params, sounds human but can't hold a conversation

alexcovo_eth · x · 2026-11-12

Developer naklecha released caramel-1, an open-weight 413M-parameter speech-to-speech model with Apache-2.0 code and a full training recipe.

Built on a Moshi-like architecture, it learns to listen and speak directly on speech without converting to text
The speech sounds remarkably human, but the model can't hold a coherent conversation yet; public demos feature two copies of the model talking to each other
Once data is prepared, training takes 2h15m (roughly 4 GPU-hours); data preparation is the larger task
The technical write-up covers architecture, data prep, training and experiments, fully reproducible

More from Multimodal