Voxtral 3 replaces Mistral's Voxtral 2 transcription and speech models and adds a voice agent SDK. Most of the upgrade is a model ID swap, but three changes can break working code: word timestamps are gone, realtime settings lock at the first audio chunk, and diarization segments can now overlap. Check those before you switch.
What was released
Voxtral 3 was announced by Mistral on 20 October 2026 in the Voxtral 3 announcement. It is a family of speech models plus an SDK for voice agents:
- Voxtral Small Transcribe 3, the higher-accuracy model for recorded audio, with diarization and context biasing across more than 50 languages.
- Voxtral Mini Transcribe 3, a smaller batch transcription model that Mistral says has 4B parameters.
- Voxtral Mini Transcribe Realtime 3, built for live agents, with native end-of-turn detection.
- Voxtral Speech 3 (model ID
voxtral-mini-tts-3-0), text to speech in batch and realtime for 35+ languages and 50+ dialects, with emotion, pace and emphasis controls and voice cloning. - A voice agent SDK built on LiveKit that runs listening, reasoning, tool calls and speech as one streaming pipeline.
Mistral says all of it is available today through its API on European, US and global endpoints, with a separate conversation for on-premises or private-cloud deployment.
What actually changed
The migration guide maps each old model to a new one: voxtral-mini-2602 to voxtral-mini-transcribe-3-0, voxtral-mini-transcribe-realtime-2602 to voxtral-mini-transcribe-realtime-3-0, voxtral-small-2507 to Voxtral Small Transcribe 3, and voxtral-mini-tts-2603 to voxtral-mini-tts-3-0. The guide says the voxtral-mini-tts-latest alias will move to Voxtral 3 within a month of general availability, so anything pinned to that alias changes voice without a code change.
The changes Mistral marks as breaking for transcription:
- No word timestamps.
timestamp_granularitiesacceptssegmentonly, and asking forwordreturns a 400 error. - Realtime settings lock early. A realtime session fixes its settings when the first audio chunk arrives, and a later
session.updatereturns an error event. - Diarization segments can overlap. Code or interfaces that assume one speaker at a time need to keep both segments.
The non-breaking changes are useful: context_bias now accepts multi-word phrases such as full product names, transcription covers more languages with automatic detection, and realtime sessions gain opt-in diarization, speaker references and turn detection. One limit to note is that context_bias entries share the context window with the audio, and an oversized list returns a 400 error.
For speech, Voxtral 2 emotion presets such as gb_jane_sad still work, but Voxtral 3 adds inline [emotion=...] tags so one voice can carry several moods. ref_audio cloning still works per request, and a cloned voice can now be saved once and reused by voice_id.
Mistral's published Voxtral 3 prices charge per minute of audio for transcription and per character for speech. Realtime transcription costs roughly twice as much per minute as batch transcription with Voxtral Mini Transcribe 3, and the Small model sits just above the Mini batch price. The voice agent API is also priced per minute, and its rate doubles when you use any reasoning model other than Mistral Small 4.
What it means for teams building voice agents
Word timestamps are the first thing to search for. Captions, transcript editors, search highlighting and compliance redaction often depend on word-level timing. If yours do, moving to Voxtral 3 is not a model ID swap: you either rebuild those features around segment timing or align words yourself after transcription. Find every timestamp_granularities call before you plan the switch.
Realtime clients need a small rewrite. Many voice clients open a session, start streaming and then adjust language or vocabulary once they know who is calling. With Voxtral 3 that later update fails. Move caller lookup ahead of the audio stream, or open a new session when the context changes.
Retest anything that reads diarized output. Overlapping segments are more accurate for real conversations, but a call summary, a CRM note or a QA scorecard that assumes one speaker at a time may merge or drop speech. Run recorded calls with crosstalk through both versions and compare what downstream steps produce.
Treat the SDK as an architecture choice, not an upgrade. The voice agent SDK puts transcription, reasoning and speech on Mistral's side, with tools that can run at Mistral or in your application. That shortens the build, but it moves call audio and tool traffic through one more provider. If you already run a voice pipeline with your own orchestration, logging and evaluation, as covered in our post on retesting a voice agent after a model switch, you can adopt the new transcription and speech models without adopting the SDK. Mistral's voice agent docs say client-side tools do not work on phone calls, and SIP telephony is offered to enterprise customers through their account team at launch.
When not to switch yet. Hold off if you depend on word timestamps and cannot rebuild that feature this sprint, or if your contract or data residency review covers Voxtral 2 model IDs only. The migration guide shows a deprecation timeline section with no dates at the time of writing, so there is no forced cut-over to plan around yet. Pin explicit model IDs rather than latest aliases until your own tests pass, and plan for retirement as described in when a model is deprecated.
Measure latency in your own stack. Mistral cites sub-second turns for the SDK and sub-500ms realtime transcription. The migration checklist itself ends with measuring latency again from your application, and that is the number your users hear. Our guide to making an AI agent respond faster covers where the rest of the delay usually sits.
Frequently asked questions
Is moving from Voxtral 2 to Voxtral 3 just a model ID change?
For many integrations, mostly yes, but not always. Mistral's migration guide lists three breaking changes: word timestamps now return a 400 error, realtime sessions reject setting changes after the first audio chunk, and diarization segments can overlap. If your code uses none of these, swapping each model ID and retesting may be enough. If it uses any of them, plan code changes first.
What replaces word timestamps in Voxtral 3?
Voxtral 3 returns segment timestamps only, so features that highlight or redact individual words need another approach. Options include keeping Voxtral 2 for that workload while it remains available, running a separate forced alignment step on the transcript, or redesigning the feature around segments. Choose based on how precise the timing must be, and test with your own recordings before committing.
Do I have to use the Voxtral voice agent SDK?
No. The transcription and speech models are available on their own through the Mistral API, so you can keep your existing orchestration, tool calling and logging. The SDK suits teams starting a new voice agent who want listening, reasoning and speech in one streaming pipeline. Teams with an established pipeline should compare its observability and data flow against what they already run.
Will my Voxtral TTS voices sound different after the upgrade?
They can. Voxtral 2 emotion presets are still accepted, but Voxtral 3 introduces new voices, inline emotion and pace tags, and an expressiveness setting. Mistral also says the voxtral-mini-tts-latest alias will point to Voxtral 3 within a month of general availability. Pin an explicit model ID and generate test samples before changing anything customers hear.
