CSM-1B

CSM-1B is the open-weights conversational speech model published by Sesame in March 2025, the company founded with Oculus co-founder Brendan Iribe. It sits underneath Maya and Miles, the hosted voice companions whose demo drew attention for sounding uncomfortably close to a person on the phone. The architecture is unusual for a speech system: a Llama-based backbone paired with a smaller audio decoder that emits Mimi audio codes, so text and audio context are consumed together rather than passed through a text-to-speech stage — that shared context is what lets the model carry conversational timing, hesitation and interruption. What matters editorially is the gap between the demo and the download. The released weights are a base generation model with no fine-tuning towards any particular voice, primarily English, explicitly published for research and education, and unable to generate text at all; the personality people reacted to lives in Sesame's hosted product, not in the checkpoint. The licence is Apache 2.0, and the card names impersonation without consent and deceptive content as prohibited uses.

▸Videos

No videos yet.

▸Reviews