American company Zyphra has released a new text-to-speech model called ZONOS3-Mini. It is designed for instant processing directly on the user's device and uses a mixture-of-experts architecture with 800 million parameters, only 100 million of which are active. This allows it to generate fluent speech in real time on phones and laptops, including running in a web browser.
Smaller and faster than its predecessor
With the new model, Zyphra is addressing the main weakness of the previous version. The ZONOS2 model had 8 billion parameters, and this size limited where it could be deployed. Training the new model did not take long—the company completed it in two and a half days using 112 Nvidia H100 graphics cards.
The speed improvement is noticeable. On a single H100 card, with eight concurrent requests, the model starts returning audio in roughly 80 milliseconds and, in the best-case scenario, can generate 457 seconds of audio in a single second. A single run can handle up to ten minutes of audio at a 44.1 kHz sampling rate.
Voice cloning and custom characters
The model can do two main things. It can imitate a voice from an audio sample or create an entirely new character based on a specified age, accent, and vocal tone. Users can control emotion, tempo, and delivery down to individual words, which is useful for games, voice-overs, and voice assistants. A separate editing model is also available, allowing sentences to be modified after they have been generated.
Czech users will appreciate its language support. ZONOS3-Mini supports 43 languages, including Czech and Slovak among its Central and Eastern European languages. For a quick start, it offers 27 preset voices grouped by language.
The test results look good, but with some caveats. In voiceprint mode, the model scored 66.2 points in the overall index, while Fish Audio S2 Pro achieved 61.7 and MOSS-TTS v1.5 scored 60.2 points. However, the test was designed and conducted by Zyphra itself. When cloning directly from a reference recording, the model scored only 44.7 points, and it trails both Fish S2 Pro and Qwen3-TTS in audio quality ratings.
Where and how to try it
The company has released the model weights under the Apache 2.0 license, so they can be downloaded and used free of charge. Those who do not want to install anything can use the hosted version in Zyphra Cloud.
The easiest way is through a browser. The model runs entirely on the user's graphics card using WebGPU, with no installation or server required. On the first visit, it is downloaded and stored by the system for future use. Developers also have access to Zyphra's own runtime options, specifically a server for Nvidia cards, a version for Apple Silicon chips, and a C++ server that runs on CPUs and GPUs via CUDA, Metal, or Vulkan. Through SGLang-Omni, the model can be deployed behind an OpenAI-compatible interface, including streaming audio playback and voice cloning without having to transcribe the sample.



