Voxtral 4B TTS 2603

Voxtral 4B TTS 2603

An open-weights text-to-speech model from Mistral with 20 preset voices in nine languages, streaming output and low latency. Made for developers who build voice agents and can accept a non-commercial license.

Made by mistralai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Voxtral TTS is an open-weights text-to-speech model from Mistral, built on Ministral-3-3B-Base-2512. It turns written text into spoken audio for voice agents. It ships with BF16 weights, 20 preset voices and the option to adapt to new voices. Audio comes out at 24 kHz in WAV, PCM, FLAC, MP3, AAC or Opus.

Who it suits: It suits developers who build real-time voice agents and want an open model they can host themselves, for non-commercial work.

Industry | Who uses it | How they use it | Example
Customer support | Contact center teams | Give a voice agent a spoken reply | A phone assistant that reads out answers
Financial services | Bank technology teams | Voice agents for identity checks | The card mentions a demo for banking KYC agents
Automotive | In-car software teams | Speak responses inside vehicles | A navigation assistant that talks back
Translation | Language service teams | Speak translated text aloud | Reading a translated message in another language

Skip it for commercial products, since the license is non-commercial, and for any use that mimics a voice you have no right to use.

Where to use it:

• vLLM-Omni, the serving library the card recommends
• A single GPU with 16 GB or more of memory
• Voice agent workflows that need streaming or batch audio

Use case | Who | What they get back
Voice agent replies | Developers | Spoken audio generated from text
Streaming speech | Real-time apps | Audio that starts quickly and plays as it arrives
Batch narration | Content teams | Many audio files from many texts
Voice customization | Developers | A new voice adapted from a short audio reference

How to set it up:

• Python, plus a GPU with 16 GB or more of memory
• vllm version 0.18.0 or newer, and vllm-omni at 0.18.0 or newer
• Disk space is not stated, so check the file listing before you download

1. Install vllm.
2. Install vllm-omni.
3. Check the mistral_common version, which should be 1.10.0 or newer.
4. Start the server with the omni flag.

pip install -U huggingface_hub
hf download mistralai/Voxtral-4B-TTS-2603

uv pip install -U vllm
uv pip install vllm-omni --upgrade

python3 -c "import mistral_common; print(mistral_common.__version__)"

vllm serve mistralai/Voxtral-4B-TTS-2603 --omni

How to use it: You send the running server a request with the text, the model name, an audio format and a voice name. The card's Python client posts to the audio speech endpoint of your server, then reads the returned audio. BASE_URL stands for your own server address.

payload = {
    "input": "Paris is a beautiful city!",
    "model": "mistralai/Voxtral-4B-TTS-2603",
    "response_format": "wav",
    "voice": "casual_male",
}

Field | What it does
input | The text to speak
model | The model name, mistralai/Voxtral-4B-TTS-2603
response_format | The audio format to return, such as wav
voice | One of the preset voices, such as casual_male

Example: You send the sentence "Paris is a beautiful city!" with the voice casual_male and the wav format. The server returns audio, and the card's client reads it and prints how many samples it got and at what sample rate.

For newcomers and experts: Think of it as a voice generator you run on your own GPU. A good first try is to start the server and send the card's sample sentence with one preset voice.

If you know the field, the card reports a real-time factor of 0.103 with 1 request at a time, using a 500-character text and a 10-second audio reference on one NVIDIA H200. At 32 concurrent requests, latency rises to 552 ms while throughput reaches 1430.78 characters per second per GPU.

Send text and a voice name
vLLM-Omni runs the model
Audio is generated at 24 kHz
Receive WAV, MP3 or another format

Title: Latency at each concurrency level
Concurrency 1: 70 ms
Concurrency 16: 331 ms
Concurrency 32: 552 ms

Good to know:

• Voices: 20 preset voices, plus adaptation to new voices.
• Languages: Arabic, Dutch, English, French, German, Hindi, Italian, Portuguese and Spanish.
• Voice rights: use only voices you have the right to use, and the card tells users to follow the law and avoid misuse.
• License: CC BY-NC 4.0, which the reference voices share, so no commercial use.
• Cost to run: none named beyond a GPU with 16 GB or more.
• Data it reaches: it downloads model files from Hugging Face, and the server handles the text and any audio reference you send it.
• Not for: commercial use, or misuse such as impersonation.

On Hugging Face, 13 November 2026

• Task: text to speech
• Library: vllm
• Downloads in the last 30 days: 2,610
• Likes: 956
• License: cc-by-nc-4.0
• Languages: en, fr, es, pt, it, nl, de, ar
• Based on: mistralai/Ministral-3-3B-Base-2512
• Quantized versions: 12
• Fine-tunes: 2
• Tags: vllm, mistral-common, text-to-speech

Links

• Model: https://huggingface.co/mistralai/Voxtral-4B-TTS-2603
• Files: https://huggingface.co/mistralai/Voxtral-4B-TTS-2603/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download mistralai/Voxtral-4B-TTS-2603

Then follow the usage in its model card. Nothing in this kit runs by itself.