Voxtral TTS: open weights under CC BY-NC versus the paid API
Mistral's Voxtral TTS has open weights under CC BY-NC 4.0 and a paid API. What the licence means for a product, and what Sume offers instead.
The Voxtral TTS weights on Hugging Face are published under CC BY-NC 4.0, and the NC means non-commercial. If you run the weights yourself inside a paid product, read the licence with a lawyer first. The commercial route is Mistral's own API at $0.016 per 1,000 characters. Sume does not host Voxtral; its TTS router serves Sonic models.
Facts from Mistral
| Item | Value |
|---|---|
| Release | 23 March 2026 |
| Languages | 9: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi, Arabic |
| Open weights licence | CC BY NC 4.0, on Hugging Face |
| API price | $0.016 per 1,000 characters |
| Latency | about 70 ms model latency for a 10 s voice sample and 500 characters |
| Voice reference | as little as 3 seconds |
Three ways to use it
- Hobby or research: download the weights and run them locally. Attribution applies.
- Product: use the API and pay $0.016 per 1,000 characters. At that rate, 50,000 characters costs $0.80.
- Mixed workflow: keep an engine you can license for commercial use and add your own steps around it.
Where Sume fits
Sume's TTS router lists sonic-3.6, sonic-3.5, sonic-3, sonic-latest and sonic-preview, at $0.0475 per 1,000 characters. That is about three times Mistral's API rate, for a different engine and a different surrounding service: job records, transcription, music generation and mixing under one key. If Hindi or Arabic is your main language, test your own script on each engine before choosing. Set language explicitly for non-English text and check the supported list in the API reference.
Limits
This post summarises a vendor announcement, not the full licence text, and it does not claim Voxtral is better or worse than Sonic. Voice cloning from a few seconds of audio raises consent questions that apply to any engine, so only clone voices you have rights to. Compare costs across vendors in the cheapest TTS API list.
Sources
More in Models
- Wan 2.2 A14B explained: 27B total, 14B active, 80GB GPU
Wan 2.2's A14B models use two experts, one for noisy early steps and one for late detail. What that means for GPU memory, and when a hosted route is simpler.
- Wan 2.2 weights: Apache 2.0, 5-second 480p/720p; Sume hosts wan-3.0
Wan 2.2 T2V-A14B is Apache 2.0 with open weights and makes 5-second 480P or 720P clips. Sume does not host Wan 2.2; its Wan row is wan-3.0, 2 to 30 seconds.
- Wan2.2 S2V-14B speech-to-video: open weights vs Sume audio options
Wan2.2-S2V-14B is an Apache 2.0 audio-driven model at 480P and 720P. On Sume: an audio reference on Wan 3.0, or a script-driven Avatar video.
- Wan 3.0 allows 20,000 prompt characters, MiniMax H3 7,000
Alibaba allows 20,000 characters in a Wan 3.0 prompt; MiniMax caps each H3 text item at 7,000. How to write one prompt that fits both on Sume.
Written by Sume