Mistral releases Voxtral TTS, a 4B open-weights speech model, non-commercial
Voxtral 4B TTS covers nine languages with 20 preset voices and streaming; weights are CC BY-NC 4.0.
Mistral AI released Voxtral 4B TTS (2603), described on its card as "a frontier, open-weights text-to-speech model" for voice agents. It supports nine languages: English, French, Spanish, German, Italian, Portuguese, Dutch, Arabic and Hindi. It ships with 20 preset voices and the card mentions "easy adaptation to new voices" through AI Studio. The card does not use the term voice cloning. Output is 24 kHz audio in several formats, with streaming and batch inference. The base model is Ministral-3-3B-Base-2512.
The card gives benchmarks on a single NVIDIA H200 with vLLM: at concurrency 1 latency is 70 ms with a real-time factor of 0.103, and at concurrency 32 it is 552 ms. It needs a single GPU with 16 GB or more, and recommends vLLM-Omni for serving.
The licence is cc-by-nc-4.0, and the reference voices are CC BY-NC 4 as well, so commercial use is not covered. The card has no release date; 26 March 2026 is the date of the linked paper on the card. The Hub repo was created earlier, in November 2025.
Open weights, yes. Open for business, no.
Sources
Editor's first pass (AI-assisted)
- Real 4: Weights are public and vLLM serving is documented; commercial use is excluded.
- Useful 3.5: Fits voice agents, but the non-commercial licence blocks many deployments.
- Notable 4: Mistral's first open TTS, with latency and throughput numbers published.
- Novel 3: A 4B model on a 3B base is conventional; the open-weights step is the news.
Comments