Futureweb

Artificial Intelligence

Mistral's Voxtral 3 is a speech-to-text model, a speech model, and a voice-agent kit. The scores are Mistral's.

Mistral's October 20 release pairs Voxtral 3 speech models with a voice-agent kit. Customers can use its API or their own machines. The quality claims are Mistral's.

The short version

Mistral's October 20, 2026 release describes Voxtral 3 as speech-to-text and text-to-speech models plus a voice-agent kit that listens, calls tools, and speaks. The docs price the text-to-speech model. The Decoder reported the launch. Quality rankings in the announcement are Mistral's.

A release dated October 20

Mistral's news post introduces Voxtral 3 as a family of speech models and a voice-agent kit for live, multi-step conversations. The Decoder's story the same day is one account. MarkTechPost's item is what the artificial-intelligence wire carried. The docs add a model card for the text-to-speech model, marked public preview and dated October 20, 2026. Two of the three sources are Mistral's. The Decoder is the other account. The document named in the sources is the record this page will keep. The document named in the sources is the record this page will keep.Exa 3 ranks its own tests.

Listen, speak, and call a tool

The post says a voice agent has to transcribe across languages and accents, decide what the person meant, speak with some control of tone, and take actions in other systems, all while the call is still going. Voxtral 3 is the company's bundle for that list. Small Transcribe 3 is the speech-to-text model, with diarization, phrase hints, and end-of-turn detection named as features. Mini TTS 3 is the speech model, with voices that the post says can change emotion and pace. The kit is what connects those models to tool calls over phone or web.what on-device inference keeps.

A price, and a deployment choice

The TTS card lists 13 dollars per million characters. It also lists a time-to-first-audio figure of about 90 milliseconds, streaming, and a set of emotion tags. Those numbers belong to the card. The news post says customers can use the EU-hosted API or run the models on their own infrastructure, and that they control whether audio is kept. A sovereign-deployment sentence is a product option. It is not an audit of a customer's logs. The document named in the sources is the record this page will keep.Clairic returns typed values.

What the announcement ranks, and who ranked it

Mistral calls the transcription model state of the art for multilingual transcription. That ranking is the company's. The Decoder does not become a second scoreboard by retelling the post. A page on this paper about evaluations a second lab can run is the standard a vendor chart still has to meet. Nothing in these sources is that rerun. Asked directly, How many languages does Mistral claim? The announcement says transcription across more than 50 languages, and speech synthesis across more than 35 languages and more than 50 dialects. Those are the company's counts.an evaluation a second lab can run.

Pro Voice is a narrower offer

The post describes a professional voice adaptation that, it says, can match a speaker from about 30 minutes of audio, across languages. That is a separate claim from the stock voices. It is still Mistral's description of its own feature. This page does not have the audio, and it does not score the match. Asked directly, Where can it run? The post says the family is available through Mistral's EU-hosted API or for deployment on the customer's own infrastructure. Self-deployment is an option the company states, not a file this page mirrors.weights you can fetch.

Preview is the doc's status

The TTS card says public preview. A preview price can change, and a preview model can move. The news post's present tense, available today, sits next to that label. Both can be true: the API is open to callers, and the card has not called the model generally available. This page keeps both sentences. Asked directly, What does the TTS doc charge? The model card dated October 20, 2026 lists 13 dollars per million characters for Voxtral Mini TTS 3, in public preview. That price is the doc's. It is not a bill this paper received.the artificial-intelligence desk.

Who is in scope

Teams building contact-center agents, assistants, or a hands-free tool on Mistral's API or on their own machines. People who only needed a transcription benchmark from a second lab are in scope as readers of a claim, not as holders of a new public score. Audio that never leaves a customer's deployment is the case the post says self-hosting is for. Asked directly, Did The Decoder rerun the models? Its October 20 story is an account of the launch. It is not a second lab's transcription score.OpenAI models enter Model Garden.

What stays on the record

October 20, 2026. Voxtral Small Transcribe 3, Voxtral Mini TTS 3, and a voice-agent kit. More than 50 languages for transcription, on Mistral's count. Thirteen dollars per million characters on the TTS card. EU API or self-deployment. The Decoder as the outside account. No second-lab score. Asked directly, What is in the family? Mistral's post describes a speech-to-text model, Voxtral Small Transcribe 3, a text-to-speech model, Voxtral Mini TTS 3, and a voice-agent kit that combines listening, reasoning, tool calls, and speaking in one streaming pipeline.Futureweb.

What would move the row

A public weight file with a license, or a second lab's transcription set, would be a later document. A price change on the card would move the thirteen-dollar line. The record today is the announcement, the card, and one retelling. The short answer this page will stand on is this: Mistral's October 20, 2026 release describes Voxtral 3 as speech-to-text and text-to-speech models plus a voice-agent kit that listens, calls tools, and speaks. The docs price the text-to-speech model. The Decoder reported the launch. Quality rankings in the announcement are Mistral's.llm-d 1.0 routes a fleet.

Sources

The reports this brief is filing. Futureweb did not republish them.

  1. Mistral, Voxtral 3
  2. Mistral docs, Voxtral TTS 3
  3. The Decoder

Questions

What is in the family?

Mistral's post describes a speech-to-text model, Voxtral Small Transcribe 3, a text-to-speech model, Voxtral Mini TTS 3, and a voice-agent kit that combines listening, reasoning, tool calls, and speaking in one streaming pipeline.

Where can it run?

The post says the family is available through Mistral's EU-hosted API or for deployment on the customer's own infrastructure. Self-deployment is an option the company states, not a file this page mirrors.

What does the TTS doc charge?

The model card dated October 20, 2026 lists 13 dollars per million characters for Voxtral Mini TTS 3, in public preview. That price is the doc's. It is not a bill this paper received.

How many languages does Mistral claim?

The announcement says transcription across more than 50 languages, and speech synthesis across more than 35 languages and more than 50 dialects. Those are the company's counts.

Did The Decoder rerun the models?

Its October 20 story is an account of the launch. It is not a second lab's transcription score.