Data read October 27, 2026 · Coval, Speko, Artificial Analysis, Cekura
Open-weights TTS models in 2026: 12 you can self-host, and how they rank
12 of the 96 models are open weights, meaning the model files can be downloaded and run on your own servers; 2 of them are in the multilingual ranking, led by Alibaba Qwen3 TTS Fast and Alibaba Qwen3 TTS 1.7b.
Speed figures come from hosted APIs measured by Coval, so a self-hosted deployment will differ with your hardware. The tag is Coval's and Artificial Analysis's, read on October 27, 2026.
How the rows are ordered. Ranked models first, in ranking order; the others follow alphabetically.
| Rank | Model | Provider | Speed tier From the Coval P95 | P95 (ms) 1 turn in 20 is slower | Languages tested |
|---|---|---|---|---|---|
| 1 = |  Alibaba Qwen3 TTS Fast | Alibaba (Qwen) | Headroom | 98.8 | 0 |
| 4 |  Alibaba Qwen3 TTS 1.7b | Alibaba (Qwen) | Fast enough | 136.6 | 0 |
| not ranked |  Boson AI Higgs Audio V3 TTS | Boson AI | n/a | n/a | 9 |
| not ranked |  BreezeBlue Breeze TTS 2 | BreezeBlue | n/a | n/a | 1 |
| not ranked |  Coqui XTTS v2 | Coqui | n/a | n/a | 7 |
| not ranked |  Fish Audio Drama 3 Preview | Fish Audio | Fast enough | 209.3 | 2 |
| not ranked |  Fish Audio OpenAudio S1 Mini | Fish Audio | n/a | n/a | 7 |
| not ranked |  Fish Audio S2 Pro | Fish Audio | n/a | n/a | 9 |
| not ranked | Guava Daytona TTS | Guava | Visible delay | 584.8 | 0 |
| not ranked |  Maya Research Maya1 | Maya Research | n/a | n/a | 0 |
| not ranked | | Mistral | n/a | n/a | 5 |
| not ranked |  Nari Qwen3-TTS | Alibaba (Qwen) | n/a | n/a | 0 |
Frequently asked questions
- What does open weights mean for a TTS model?
- The trained model is published so you can download and run it yourself. It does not always mean the code or the training data are open, and licences differ.
- Which open-weights TTS models are fastest?
- Starting with the best-ranked, Alibaba Qwen3 TTS Fast and Alibaba Qwen3 TTS 1.7b lead the ranked open-weights models. Speed was measured on hosted endpoints.
- Can I self-host a multilingual real-time TTS model?
- Yes, for the open-weights models above. Check each licence and run your own latency test on your hardware.