Voxtral 3: Mistral Expands Language Models to 35+ Languages

TVS News AI, editorially supervised, presents Voxtral 3: Mistral Expands Language Models to 35+ Languages.

Mistral AI is positioning Voxtral 3 directly alongside ElevenLabs. According to Frandroid, the new model family transcribes more than 50 languages and speaks in more than 35. If you are comparing multilingual AI voices or voice agents, these are the numbers you can initially work with.

According to the report, Voxtral 3 is not a single voice but a family of language models. The models can transcribe speech, speak in more than 35 languages, and form the technical foundation for a toolkit for building speech-based agents. Mistral is thus bundling several capabilities into a single product line.

The biggest leap has also come in the few months since March: Mistral's first speech-synthesis model could support only 9 languages at the time. Voxtral 3 now speaks in 35+ languages. For context, that means Mistral has significantly expanded its multilingual capabilities, while the new family also covers transcription in 50+ languages.

Voxtral 3 Transcribes 50+ Languages and Speaks in 35+

Two capabilities, two sets of numbers

With Voxtral 3, the language figures apply to different functions: more than 50 languages refers to transcription, while more than 35 languages refers to speech. This distinction is important when comparing the new model family with ElevenLabs.

For ElevenLabs, Frandroid gives a blanket figure of 90 languages. By contrast, the report cites two separate figures for Voxtral 3, covering two separate tasks. It would therefore be too crude to force all the numbers into a single ranking without making this distinction. What is clearly established is that Mistral offers broad language coverage in both transcription and speech output.

From 9 to More Than 35 Languages Since March

Significantly more multilingualism in a few months

In March, Mistral's first speech-synthesis model supported only 9 languages. Voxtral 3 now supports more than 35 languages for speech output. It can also transcribe in more than 50 languages. The new model family therefore offers broader language coverage for both tasks.

The figures cited by Frandroid do not reveal how well the additional languages perform in practical use. The expansion of language coverage is therefore clearly measurable, while claims about the quality of individual languages would require separate testing.

Voxtral 3 Will Serve as the Foundation for an Agent Toolkit

Not just speech, but a development path

According to Frandroid, Voxtral 3 is intended not only to transcribe or output speech. The model family also serves as the foundation for a toolkit for building speech-based agents. As a result, the focus is not solely on the number of supported languages, but also on how speech capabilities can be used to develop complete agents.

The report provides no specific details about the toolkit: it mentions neither a dedicated product name, prices, nor any standalone availability information. The only firmly established fact is its role as a technical foundation. That is precisely where the toolkit offers a practical addition to the language figures.

What the Comparison With ElevenLabs Really Shows

Choose by function, not just by counting languages

For your comparison, the specific features are worth considering first: 50+ transcribed languages, 35+ spoken languages, and a foundation for speech-based agents. By contrast, ElevenLabs' cited 90 languages represent a single overall figure. Which offering is better suited to your use case therefore depends on which of these features you actually need.

The report provides no concrete prices, quality comparisons, or details about the availability of the agent toolkit. Further information would therefore be needed for a reliable decision. The current state of Voxtral 3 can nevertheless be clearly summarized: from 9 to 35+ spoken languages, along with 50+ transcribed languages and a toolkit that provides a development foundation.