San Francisco speech-model company Zyphra open-sourced ZONOS3-Mini on November 9. It is a text-to-speech model built on a sparse mixture-of-experts (MoE) architecture, with 800 million total parameters of which only 100 million are activated per inference. It ships under the Apache 2.0 license, so commercial use is allowed. The weights are on Hugging Face and GitHub, Zyphra's own cloud service went live at the same time, and the company's website has a page where you can try the model directly in a local browser.
The previous generation, ZONOS2, had 8 billion parameters and a codec frame rate of 86 Hz. This time the parameter count drops by an order of magnitude, and the team has put its emphasis on where the model can be deployed:
"we release runtimes for browsers via WebGPU, servers via SGLang and vLLM, as well as many others through GGML"
Latency and training scale
Zyphra gives two latency figures. In the browser on a Pixel 9 phone, the first chunk of audio arrives in under 570 milliseconds after text is entered; on a MacBook Air with an M3 chip it takes under 150 milliseconds. Zyphra says the model can run in real time on phones several years old, but it has not published a specific real-time factor (RTF).
Training took 112 Nvidia H100 GPUs running for 2.5 days. On the server side, SGLang and vLLM support was available on release day, and other devices can use GGML's native C++ runtime.
What you can control
Voice cloning is zero-shot and comes in two forms. In the first, a separate audio clip is placed in the context during pretraining, which yields a more natural voice but does not reproduce the original speaker's prosody exactly. In the second, the model uses the voice representation of the target audio directly and then runs a brief fine-tune. The speaker encoder has been switched from the Qwen3-TTS approach to MECT-B2.
The adjustable dimensions fall into three layers. The identity layer covers age, gender, accent and voice quality; the expression layer covers speaking rate, emotion, projection and liveliness; and going finer, alignment attention heads allow word-level control. The team also released a VoiceCraft-based speech editing model for changing parts of an existing recording.
The ranking comes from the company's own benchmark
Zyphra evaluated the model with its self-built ZTTS2-Eval, whose clean set covers 35 languages and whose in-the-wild set covers 20 languages. The composite Zyphra TTS Capability Index v1 is made up of 420 items. On that index, ZONOS3-Mini ranks first among open-source models, followed by Fish S2 Pro and MOSS TTS v1.5.
The release page shows only a Pareto curve of latency versus throughput and does not list per-model scores. It also does not spell out which languages ZONOS3-Mini itself supports. The two projects ranked second and third have so far not responded to the ranking.
What it means for Chinese-language developers
The Fish and MOSS families both have a sizable user base in China's open-source speech synthesis community, and both appear on ZONOS3-Mini's comparison list. However, although the evaluation spans 35 languages, no separate score for Chinese has been published, so how it handles Mandarin polyphonic characters, erhua (the retroflex r-suffix) and Cantonese can only be judged by downloading the model and trying it.
Because the browser path runs on WebGPU, use cases such as web page read-aloud and voiceover for online courseware can avoid sending user text to a server, which makes things easier for projects with data-compliance requirements. Access to Hugging Face from mainland China is unreliable, so downloading the weights may require a mirror site. Apache 2.0 removes the hassle of commercial licensing, but voice cloning and speech editing count as deep synthesis services under Chinese rules, so before launch you still need to obtain consent from the person whose voice is cloned and add prominent labeling, as the relevant regulations require.
Sources: Zyphra's official model release page, CocoLoop; checked parameter scale, license, first-audio latency on browser and laptop, training compute and benchmark ranking.