Made by AI, Directed by David Stein

Search
AI & Creative Tools

Mistral's Voxtral 3 Is Very Good and, For the First Time, You Cannot Download It

Three transcription models, a TTS model and a LiveKit-based voice agent SDK — all in public preview on the API, with no weights on Hugging Face.

5 min read

Mistral announced Voxtral 3 on 20 October 2026: a family of speech models plus a voice agent SDK. Three transcription models, one text-to-speech model, one SDK built on LiveKit, 50+ languages. On the benchmarks it is genuinely strong.

It is also the first Voxtral release you cannot run from downloaded weights, and almost nothing written about it has said so.

The thing to check before you plan around it

All four models are in public preview on the Mistral API. The announcement offers “Mistral’s EU-hosted API or for self-deployment on your own infrastructure” — but self-deployment routes through Mistral sales, not through a download. GPU and VRAM requirements are not disclosed. The size of Voxtral Small Transcribe 3 is not disclosed.

We checked the Hugging Face API for what Mistral actually publishes under the Voxtral name. There are five mistralai/ Voxtral repositories, and every one of them predates Voxtral 3:

RepositoryCreatedDownloadsLicence
mistralai/Voxtral-Mini-3B-25072025-07-01142,198Apache-2.0
mistralai/Voxtral-Small-24B-25072025-07-0118,479—
mistralai/Voxtral-Mini-4B-Realtime-26022026-01-212,154,081Apache-2.0
mistralai/Voxtral-4B-TTS-26032025-11-171,575—
mistralai/Voxtral-Mini-4B-Realtime-Arabic2026-10-081,518no licence tag

Searching every Voxtral repository on the Hub, exactly one has been created since 19 October, and it is a community GGUF conversion of the older Arabic realtime model, not of anything in Voxtral 3.

So the timeline is: Mistral shipped Apache-2.0 Voxtral weights in July 2025, shipped the realtime model in January 2026 that has since been downloaded 2.15 million times, shipped another open checkpoint on 8 October 2026 — and then, twelve days later, shipped the next generation as an API preview. The word “weights” does not appear in the announcement. Neither does “open.”

That may be temporary. Mistral has previously opened weights after an API period. But a studio planning a local pipeline around Voxtral 3 today is planning around a product that does not exist in that form, and the coverage repeating the press release will not tell you that.

What is in the family

  • Voxtral Small Transcribe 3 — flagship batch transcription with diarization. The docs also list translation, speech Q&A, summarisation and function calling.
  • Voxtral Mini Transcribe 3 — 4B parameters, for cheaper and faster batch jobs. Mistral says Mini suits resource-constrained environments.
  • Voxtral Mini Transcribe Realtime 3 — streaming ASR over WebSocket with end-of-turn detection.
  • Voxtral Speech 3 / Voxtral Mini TTS 3 — text-to-speech, 35+ languages and 50+ dialects, with voices that adjust emotion, pace and emphasis from conversational context.
  • Voice agent SDK — built on LiveKit, combining listening, reasoning, tool calling and speaking in one real-time streaming pipeline.

Both batch transcription models list a 32k context.

The numbers, including the bad one

  • 2.1% AA-WER for Voxtral Small Transcribe 3, placing it #3 on Artificial Analysis’ non-streaming speech-to-text ranking. Mistral describes it as strongest on diarization and multilingual batch transcription.
  • At the other end: the realtime model at 240 ms latency scores 7.4% WER on FLEURS 35, against 4.2% for Gemini 3.5 Realtime.

That second figure is the useful one and it is in the write-ups to Mistral’s credit. Batch transcription is state-of-the-art adjacent; streaming at low latency is well behind the best. If your use case is subtitling a finished recording, this is a top-three option. If it is live captioning on stage, the gap is nearly two to one in error rate.

Pro Voice: thirty minutes

The headline feature for anyone doing narration or dubbing is Pro Voice, pitched for “professional use cases where basic voice cloning does not suffice.” Mistral says it reproduces a speaker’s distinctive character, emotional range and speaking style across languages, from 30 minutes of audio.

Thirty minutes is not much. Any podcast guest, any recorded lecture, any artist with a back catalogue of talks has handed over thirty minutes of clean audio many times. The cross-language claim makes it more pointed: the output is a voice saying things in languages its owner does not speak. We have written before about where expressive voice cloning leaves consent, and nothing about a 30-minute threshold on a sovereign-hosted API changes the shape of that problem.

The sovereignty pitch

Mistral’s framing throughout is data sovereignty: EU-hosted endpoints, control over where audio is processed, control over whether conversations are retained. For a European broadcaster, clinic or public body that is a real differentiator and the reason this release exists in the form it does.

It is worth being precise about what that pitch is and is not. Sovereign hosting answers whose jurisdiction processes your audio. Open weights answer whether you can run it at all without us. Voxtral 3 improves the first and, for now, removes the second. Those are different guarantees, and the announcement’s structure encourages reading the first as if it delivered the second.

Verdict

As a product: strong batch transcription, a credible TTS with contextual expressivity, an SDK that puts the whole loop in one pipeline, and an honest published weakness in streaming latency. If you are building a voice agent on an API, this is a serious option and the EU hosting may be decisive.

As a release: check the download page before you architect. For a local speech pipeline today, the Apache-2.0 Voxtral-Mini-4B-Realtime-2602 is still there with 2.15 million downloads behind it, and it is not going anywhere. Voxtral 3 is not that, yet.

Sources

Related Reading