Voxtral TTS
Mistral's 4B open-weights TTS model: 9 languages and voice adaptation from a short clip, under a non-commercial CC BY-NC license.
In short. Voxtral TTS reads text aloud in a natural voice in nine languages, and it can adapt to a voice from a short sample. It is Mistral's first text-to-speech model, and the weights are open. The license is non-commercial, so it suits research and personal projects, while a business needs Mistral's hosted service or a different license.
Summary
Mistral released Voxtral TTS on 23 March 2026, calling it its first text-to-speech model. The weights are on Hugging Face as Voxtral-4B-TTS-2603, and a hosted version runs in Mistral's AI Studio. The model has about 4B parameters in three parts: a 3.4B transformer decoder backbone built on Ministral 3 3B, a 390M flow-matching acoustic transformer and a 300M neural audio codec. For each audio frame the backbone predicts a semantic token, and the flow-matching transformer then runs 16 function evaluations to produce the acoustic detail. The codec works at 12.5 frames per second.
It speaks nine languages (English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi and Arabic), ships with 20 preset voices, writes 24 kHz audio in WAV, PCM, FLAC, MP3, AAC or Opus, and supports streaming. It adapts to a custom voice from a short reference clip. Mistral says it can also carry a voice across languages, such as English speech with a French voice prompt, although it was not trained for that. Mistral reports a model latency of 70 ms for a 10-second voice sample and 500 characters, and a real-time factor of about 9.7. Its human-preference comparisons against ElevenLabs are Mistral's own evaluations.
The weights are licensed CC BY-NC 4.0. The model card explains that the reference voices are licensed CC BY-NC 4.0 and that the model inherits that license, so commercial use is not allowed. This differs from Mistral's other open speech model, Voxtral Realtime, which is Apache 2.0. For a speech generator that you can use commercially, look at Qwen3-TTS or VoxCPM2, both Apache 2.0.
Good for
- Voice-agent prototypes that need low-latency, expressive speech in nine languages
- Research and personal projects that adapt a voice from a short reference clip
- Speech-in, speech-out pipelines together with Voxtral Realtime
Watch out for
- The CC BY-NC 4.0 license, inherited from the reference voices, rules out commercial products unless you use Mistral's hosted service or get a different license
- Only nine languages, with no Chinese, Japanese or Korean
- The comparisons with ElevenLabs are human evaluations run by Mistral
- Copying a real person's voice needs their consent, and the watermarking or misuse safeguards of your own product are your responsibility
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Voxtral TTS. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Voxtral 4B TTS BF16 | 4.09B | 8 GB | 9.2 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
A Creative Commons license that lets anyone copy, share and adapt the work with credit, but only for non-commercial purposes. Some model cards name it without a version number, and the current version is 4.0.
- Attribution or model name must be shown
See also
- Open-weight AI models made in Europe
- Best open-source text-to-speech models in 2026
- Open speech models: transcription is open, text-to-speech is mixed
Frequently asked questions
- What is Voxtral TTS?
- Voxtral TTS reads text aloud in a natural voice in nine languages, and it can adapt to a voice from a short sample. It is Mistral's first text-to-speech model, and the weights are open. The license is non-commercial, so it suits research and personal projects, while a business needs Mistral's hosted service or a different license.
- Can I use Voxtral TTS commercially?
- Not without separate permission. Voxtral TTS is released under the CC BY-NC, which allows research and personal use only.
- Is Voxtral TTS open source?
- Voxtral TTS is open-weight, not open source in the strict sense. You can download the model files, but the CC BY-NC adds conditions that the Open Source Initiative definition does not allow.
- Can I run Voxtral TTS on my own computer?
- Yes. The smallest variant, Voxtral 4B TTS, needs about 9.2 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Voxtral TTS released?
- Voxtral TTS was released by Mistral AI on March 23, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Mistral AI: Voxtral TTS (23 Mar 2026) mistral.ai
- Voxtral-4B-TTS-2603 model card (Hugging Face) huggingface.co
- Voxtral TTS research paper (arXiv 2603.25551) arxiv.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3-TTS Alibaba Qwen | 910M – 1.92B | – | Apache 2.0 (Permissive) | 3.68 GB Laptop or phone | |
| Pocket TTS Kyutai | 18M – 100M | – | CC BY 4.0 (Permissive) | 0.127 GB Laptop or phone | |
| Kokoro hexgrad | 82M | – | Apache 2.0 (Permissive) | 0.38 GB Laptop or phone | |
| Voxtral Realtime Mistral AI | 4.43B | 128K | Apache 2.0 (Permissive) | 10 GB Laptop or phone |