On October 20, 2026, Mistral released the Voxtral 3 speech family and Voice Agents SDK. The launch brings recorded and live transcription, speech generation, and tool-connected conversations into one product family. Its four speech models remain public previews, and their language coverage and deployment terms differ.
Four speech models, not one universal voice model
The launch separates recorded-audio transcription, live listening and speech generation. Mistral’s speech-to-text documentation recommends starting with Mini for recorded files and moving to Small when custom vocabulary does not resolve remaining errors. Realtime is a separate model for live sessions, not simply a streamed response from a file-transcription request.
- Voxtral Small Transcribe 3, identified as voxtral-small-transcribe-3-0, is the flagship offline transcription and diarization option.
- Voxtral Mini Transcribe 3, voxtral-mini-transcribe-3-0, is the smaller recorded-file option. Both offline variants support timestamps and speaker labeling.
- Voxtral Mini Transcribe Realtime 3, voxtral-mini-transcribe-realtime-3-0, handles streaming transcription.
- Voxtral Mini TTS 3, listed in the documentation as Voxtral TTS 3 and identified as voxtral-mini-tts-3-0, generates spoken output.
Those are the four models named in the October 20 release entry; they should not be confused with the separate reasoning model used inside a voice agent.
Language coverage differs materially. The current language specification lists 50 languages for each offline transcription model, 45 languages and dialects for Realtime, and 35 languages with 50-plus dialects for TTS. The announcement uses broader, inconsistent Realtime wording, including “over 35” and “40+”; use the model-specific specification when scoping a pilot. A transcription language count does not establish equivalent spoken-response coverage.
The SDK connects speech to business actions
Built on LiveKit, the Voice Agents SDK connects audio capture, transcription, a reasoning model, speech generation and playback. Python and TypeScript bindings support application integration, and callers can interrupt spoken replies. The business application still defines knowledge, tools and end-of-call behavior. This is an orchestration layer, not evidence that transcription alone can safely execute a workflow.
One consequential restriction: SIP telephony is available to enterprise customers through their account team at launch. Mistral also warns that client-side tools fail on phone calls; telephone agents need server-side tools. A browser proof of concept therefore does not establish that the same tool wiring will work on a support number.
API access is not an open-weight release
The October 20 announcement says API access is available and identifies European, US and global endpoints. It also offers on-premises or private-cloud deployment, but directs customers to contact Mistral for the full stack. The four model cards label these releases Premier public previews.
As of this November 1 review, those primary release pages do not supply Voxtral 3 weight-download links or an open-weight license. Treat self-deployment as a vendor-arranged option whose terms need confirmation—not permission to download and run every variant commercially. Likewise, an SDK running on a company laptop does not mean inference is local: the documented managed pipeline runs models at Mistral.
Launch restrictions change what a pilot can prove
The detailed documentation narrows several headline capabilities. Realtime diarization and end-of-turn detection cannot share one session. Offline transcription provides segment timestamps, not word timestamps. TTS supports emotion and pace tags, but free-form style instructions are marked unavailable at launch. These distinctions matter for speaker-aware call automation, transcript alignment and brand-voice requirements.
Our operational assessment: split evaluation into recorded-call processing and live task completion. For a booking agent, test corrected reference numbers, interruptions during confirmation, slow backend responses and failed writes. Require an explicit success response from the booking system before the agent says a reservation exists. These are proposed acceptance tests, not results from firsthand Voxtral testing.
The next useful milestones are clearer self-deployment terms, preview-to-GA changes and resolution of documented feature restrictions. For now, Voxtral 3 creates a credible pilot opportunity—not a reason to assume every advertised capability is simultaneously available in every channel.