Overview
Voxtral TTS is the text-to-speech model Mistral announced on March 23, 2026, completing its Voxtral speech line with a model that speaks as well as listens. It was released on the Mistral API as voxtral-mini-tts-latest through the /v1/audio/speech endpoint, in Mistral Studio, and as open weights on Hugging Face (mistralai/Voxtral-4B-TTS-2603) under the CC BY-NC 4.0 non-commercial licence.
The 4B-parameter model has three parts: a 3.4B transformer decoder backbone built on Ministral 3B, a 390M flow-matching acoustic transformer, and a 300M neural audio codec. It speaks 9 languages — English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi and Arabic — and adapts to a custom voice from a reference as short as 3 seconds. Mistral quotes a model latency of 70 ms for a 10-second voice sample and 500 characters, a real-time factor of about 9.7x, and up to two minutes of audio generated natively.
In Mistral's side-by-side human evaluations across the 9 languages, the company reports better naturalness than ElevenLabs Flash v2.5 at a similar time-to-first-audio and parity with ElevenLabs v3. API pricing was $0.016 per 1,000 characters. The research paper is on arXiv (2603.25551). Voxtral TTS 3 followed in October 2026 as a Premier, closed-weight successor.
| Released | 2026-03-23 |
|---|---|
| License | CC BY-NC 4.0 |
| Weights | Open weights |
| Parameters | 4B (3.4B decoder + 390M flow-matching acoustic transformer + 300M codec) |
| Architecture | Transformer decoder backbone built on Ministral 3B, a flow-matching acoustic transformer and a neural audio codec |
| Modalities | Text, Audio |
| Status | Released March 23, 2026 on the Mistral API and as open weights on Hugging Face. |
Pricing
| Input | $0.016 / 1K characters |
|---|
Mistral API price at launch, March 2026.
Strengths
- Open weights on Hugging Face (CC BY-NC 4.0), runnable with vLLM
- 4B parameters with 70 ms model latency and a real-time factor of about 9.7x
- Voice adaptation from a reference as short as 3 seconds
- 9 languages, including Hindi and Arabic
- API price of $0.016 per 1,000 characters
Best for
- Reach for it to run text-to-speech on your own hardware for research or other non-commercial use, where the CC BY-NC 4.0 licence fits.
- Reach for it for low-latency voice output in prototypes of voice agents across its 9 languages.
- Use Voxtral TTS 3 instead for professional voice cloning, emotion tags or commercial self-hosting arranged with Mistral.
How to access
| Provider | Model ID |
|---|---|
| Mistral API ↗ | voxtral-mini-tts-latest |
| Hugging Face (open weights) ↗ | mistralai/Voxtral-4B-TTS-2603 |
Voxtral TTS (text-to-speech) — every version
The full lineage of the Voxtral TTS (text-to-speech) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Voxtral TTS 3current | 2026-10-20 | — | Proprietary |
| Voxtral TTS | 2026-03-23 | — | CC BY-NC 4.0 |
FAQ
Can I use Voxtral TTS commercially?
The open weights are released under CC BY-NC 4.0, a non-commercial licence. Commercial use goes through the Mistral API, which listed the model at $0.016 per 1,000 characters.
Which languages does Voxtral TTS speak?
Nine: English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi and Arabic.
How much audio does it need to clone a voice?
Mistral says it adapts to a custom voice from a reference as short as 3 seconds; its architecture notes describe voice prompts of 5 to 25 seconds.