Voxtral 3: Mistral's new speech-to-text, text-to-speech and voice-agent SDK are API-only, unlike earlier Voxtral models

Voxtral 3 is Mistral AI's new speech family, released on October 20, 2026: three transcription models, a text-to-speech (TTS) model and a Voice Agents SDK that joins them to a Mistral language model. The headline for anyone tracking open models is what's missing. Every earlier Voxtral release included at least one open-weight model, but none of the Voxtral 3 models has downloadable weights. They are paid API previews, with self-hosting available only through Mistral's sales team.

This story is based on Mistral's Voxtral 3 announcement, its model and SDK documentation, Mistral's Hugging Face pages and Artificial Analysis' speech-to-text leaderboard, checked October 23, 2026. Mistral's most recent open-weight release is the Timestral forecasting family.

The short version

• Transcription: Voxtral Small Transcribe 3 ($0.004 a minute), Mini Transcribe 3 (4B parameters, $0.0035) and Mini Transcribe Realtime 3 ($0.008).
• Speech: Voxtral TTS 3, $13 per million characters, with voice cloning.
• Voice agents: an SDK built on LiveKit, $0.04 a minute with Mistral Small 4 or $0.08 with a larger model.
• Weights: none. The docs label every model "Premier" and "Public Preview"; Hugging Face has no Voxtral 3 repository.
• Prices rose: the batch Mini model costs 17% more per minute than Voxtral Mini Transcribe 2, and the realtime model 33% more than February's Voxtral Realtime.

Is Voxtral 3 open source?

No. Mistral's announcement doesn't mention weights or a license. Its docs tag all four models "Premier" (Mistral's docs tag Mistral Large 4, by contrast, "Open"), and none appears in the Voxtral collection on Hugging Face. The announcement says the models are "available today through Mistral's EU-hosted API or for self-deployment on your own infrastructure," but self-deployment goes through a contact form.

That's a change. Voxtral Mini 3B and Voxtral Small 24B (July 2025) and Voxtral Realtime 4B (February 2026) were released under Apache 2.0. The March 2026 Voxtral 4B TTS model came under CC BY-NC 4.0, which bars commercial use. An Arabic fine-tune of the realtime model, Voxtral-Mini-4B-Realtime-Arabic, went up on Hugging Face in October 2026, also under Apache 2.0. Commenters in the Hacker News discussion noticed: "Too bad it's not open, unlike the previous version," one wrote (Tier 3, community reaction). Our open-weight licenses guide explains what those licenses allow.

What's in Voxtral 3, and what does it cost?

Last verified: October 23, 2026

Small Transcribe 3's model page says it also supports translation and speech Q&A, but Mistral's audio overview describes it as "Transcription only at launch." Mistral's audio docs also list a separate "Voxtral TTS Realtime 3." The model IDs are voxtral-small-transcribe-3-0, voxtral-mini-transcribe-3-0, voxtral-mini-transcribe-realtime-3-0 and voxtral-mini-tts-3-0.

How many languages does Voxtral 3 support?

Mistral says batch transcription covers "50+ languages," including the 24 official EU languages, and its benchmark chart lists 45. Realtime transcription is benchmarked on 35. For TTS, the announcement says "35+ languages and 50+ dialects," but the TTS 3 model page lists 9 languages. Mistral hasn't reconciled the two, so check the docs for your language before you build on it.

How fast is it?

These latency figures are vendor-reported:

• Realtime transcription: "sub-500ms latency," with a delay you can set from 80 ms to 2,400 ms.
• TTS 3: about 90 ms to the first audio when streaming.
• Voice agents: "each turn completes in under one second," because, Mistral says, "no stage waits for the one before it."

On accuracy, Mistral says Small Transcribe 3 beats Whisper, Scribe v2 and Gemini 3.5 Transcribe on its multilingual FLEURS test, and its docs say Mini Transcribe 3 ranks third on FLEURS and English diarization. These are Mistral's own evaluations.

The first outside results come from Artificial Analysis. Its speech-to-text leaderboard, checked October 23, puts Small Transcribe 3 third on its AA-WER index, with 2.1% of words transcribed incorrectly, behind Alibaba's Fun-Realtime-ASR-preview (1.7%) and Microsoft's MAI-Transcribe-2 (2.0%). Mini Transcribe 3 scores 2.2%, level with ElevenLabs' Scribe v2. The index averages error rates over about eight hours of audio from three datasets, so it isn't comparable with Mistral's FLEURS figures, and the page doesn't date the runs. It doesn't list the realtime model or TTS 3.

How does the voice-agent SDK work?

The Voice Agents SDK "builds on LiveKit for the real-time audio transport," the docs say, with Python and TypeScript bindings. Each turn runs speech recognition, a Mistral language model and speech synthesis as one streaming pipeline, and you can swap the model, voice or language per agent. It supports function calling and retrieval over your own systems. Phone calls over SIP are "available to enterprise customers at launch, through your account team." The models run on Mistral's servers, on European, US or global endpoints.

The $0.08-a-minute tier matches what xAI charges for Grok Voice 3.0, covered in our voice-model story.

Is Voxtral 3 the same as Voxtral Mini 3B?

No. Searches for "voxtral 3" mostly return "Voxtral Mini 3B," which is a different, older model. In that name "3B" means 3 billion parameters, not generation 3. Voxtral Mini 3B (Voxtral-Mini-3B-2507) came out in July 2025 under Apache 2.0, and its weights are still on Hugging Face. It is a speech-understanding model built on Ministral 3B and supports eight languages. Voxtral 3 is the October 2026 API family. If you need a speech model you can download, the Apache 2.0 Voxtral Realtime 4B and Google's EmbeddingGemma 2 (in our speech-to-text roundup) are options.

What's confirmed, reported and uncertain

• Confirmed (Mistral docs): model IDs, prices, "Premier" and "Public Preview" labels, release date October 20, the 4B size of Mini Transcribe 3, the LiveKit-based SDK.
• Confirmed (Hugging Face): no Voxtral 3 repositories; earlier Voxtral licenses.
• Vendor-reported: latency, language counts and Mistral's own accuracy rankings.
• Independent: Artificial Analysis' AA-WER results for the two batch transcription models.
• Uncertain: whether Mistral will ever release Voxtral 3 weights; the TTS language count (9 or 35+); sizes of the other models.