← kwindla hultman kramer

Marvis, a new open source, local-first text-to-speech model from @Prince_Canuma…

August 26, 2025

Marvis, a new open source, local-first text-to-speech model from @Prince_Canuma and @lllucas.

I dropped it into the macos-local-voice-agents repo to play around. Really nice initial release! The model leverages architecture elements from Sesame and Moshi that enable low-latency streaming. This has been a missing piece for local TTS, so I'm excited about working more on the Pipecat integration.

Launch announcement from @Prince_Canuma[1]

Hack on the macos-local-voice-agents tools with us[2]

  1. https://x.com/Prince_Canuma/status/1960399829290426448
  2. https://github.com/kwindla/macos-local-voice-agents/