Zonos v0.1

Zonos-v0.1 is the open-weight text-to-speech model Zyphra released in February 2025, trained on more than 200,000 hours of multilingual speech (majority English, with substantial Chinese, Japanese, French, Spanish, and German). It ships in two 1.6B-parameter flavors — a pure transformer and a hybrid transformer/Mamba backbone, the latter claimed as the first TTS model to use the Mamba state-space architecture. Conditioning supports fine control over speaking rate, pitch variation, audio quality, and emotions like happiness, fear, sadness, and anger, with native 44kHz output.

Its headline feature is voice cloning from as little as 5–30 seconds of reference audio, and unusually for a model of this quality, the weights are released under the permissive Apache 2.0 license. The honest picture: users report occasional audio artifacts and alignment issues, real-time-factor performance realistically wants an RTX 4090-class GPU, and the same few-second cloning that makes it powerful is exactly what makes it a deepfake risk — Zyphra's own framing acknowledges responsible-use expectations.