Summary:
ZONOS3-Mini targets on-device speech generation. A matched English benchmark slice shows stronger speaker similarity alongside trade-offs in word error rate and predicted quality.Zyphra’s ZONOS3-Mini brings voice-cloning text-to-speech to local devices and browsers. Announced November 9, the model is available through downloadable weights, open runtimes and Zyphra’s hosted inference service.
The sparse mixture-of-experts model has 800 million total parameters, with 100 million active. Zyphra reports time to first audio below 570 milliseconds in a Pixel 9 browser and below 150 milliseconds on an M3 MacBook Air. These are device-specific first-audio measurements, not end-to-end timing for every utterance.
A matched three-model scorecard
The following developer-reported ZTTS2-Eval v1 results use the English Clean set, 1,279 prompts per model. All three configurations clone through speaker embeddings without a reference transcript. Keeping that conditioning class fixed avoids mixing transcript-assisted systems into the same comparison.
| Metric | ZONOS3-Mini embedding | Qwen3-TTS 0.6B | Qwen3-TTS 1.7B |
|---|---|---|---|
| Clipped word error rate ↓ | 2.340% | 2.198% | 1.886% |
| Speaker similarity ↑ | 80.89 | 70.08 | 68.47 |
| MSR-UTMOS predicted quality ↑ | 3.4689 | 3.8874 | 3.8715 |
| Outputs above 50% WER ↓ | 0.000% | 0.078% | 0.000% |
Bold marks each row’s best numeric value, including ties. Speaker similarity is a cosine score multiplied by 100; MSR-UTMOS predicts quality rather than measuring listener preference. English WER uses Whisper-large-v3, with each utterance’s error rate capped at 100% before averaging.
ZONOS3-Mini leads this slice on speaker similarity, while the Qwen variants lead on the other non-tied metrics. The WER confidence intervals overlap: 2.024–2.667% for ZONOS3-Mini, 1.900–2.513% for Qwen 0.6B and 1.616–2.167% for Qwen 1.7B. This is one language and recording condition, not an overall multilingual ranking. Sampling configurations also differ: ZONOS uses a fixed seed, while the Qwen runs use unseeded defaults.
Access and deployment
The Hugging Face model card lists Apache 2.0 for the weights. The separate inference-runtime repository is MIT-licensed. It offers a browser WebGPU build, HTTP servers and an offline Python API, so developers can choose between local operation and hosted infrastructure.
Delivery controls cover areas such as pace, mood and accent, including word-level steering. For an application, test the intended language, device and voice with consented recordings before committing to a deployment route. The official benchmark’s listening samples can complement the numerical comparison.
Speech generation is distinct from transcription. For the reverse workflow, see our Scribe v3 and v3 Live speech-to-text comparison.
Sources and artwork
- Zyphra launch announcement, November 9, 2026.
- Official model weights and license; inference runtimes.
- ZTTS2-Eval explorer and version 1 results and conditioning details, retrieved November 12, 2026.
- Original chart from Zyphra’s published results. The official Zyphra provider wordmark is used under its press-and-benchmark guidance; it identifies the provider rather than a distinct model logo.
