Mistral Voxtral 3 brings transcription, speech generation and tool-connected voice agents into one product family. For developers, the important launch detail is not simply a more natural voice: it is a documented path from live audio to reasoning, business tools and a spoken reply—with migration constraints worth checking before production.
Mistral’s October 20 announcement introduces new speech models alongside a LiveKit-based voice-agent SDK. This guide separates the launch claims from the current documentation and explains what I would validate in a practical pilot. FindMilan has not independently benchmarked these models.
What changed with Voxtral 3?
The release covers three related jobs: transcribing recordings, recognizing live speech and generating spoken output. The speech-to-text documentation distinguishes offline Mini and Small models from the realtime Mini model. An application that only needs a meeting transcript does not need an entire conversational agent.
The voice-agent SDK connects realtime transcription, a general reasoning model, tool calls and speech output. Its stages remain separate, so a developer can configure the reasoning model and voice instead of treating the stack as one indivisible model.
Mistral reports broad multilingual coverage and strong benchmark results. Coverage differs by mode: do not assume every transcription language also has a matching preset voice, dialect or realtime capability. Test the exact language, accent and vocabulary your customers use.
What does the API cost at launch?
The announcement lists these component rates in US dollars:
| Component | Published launch rate |
|---|---|
| Mini offline transcription | $0.0035 per minute |
| Mini realtime transcription | $0.008 per minute |
| Small transcription | $0.004 per minute |
| Speech generation | $13 per million characters |
These are not an all-inclusive price per voice-agent call. Reasoning, tool execution, telephony and application infrastructure can contribute to the total. As simple arithmetic, 1,000 minutes at the listed Mini realtime transcription rate is $8 for that component alone—not a quote for a complete contact-centre deployment.
Measure costs against completed tasks and failed or escalated calls, not just audio minutes. A cheap conversation that books the wrong appointment can be expensive operationally.
Browser access and telephone access are different
The voice-agent documentation supports Python and Node.js SDK applications and a browser TypeScript path. For a browser application, your server creates the session so a long-lived API key never reaches the page.
SIP telephone access is enterprise-only at launch, arranged through an account team. A phone call does not run your browser application, so client-side functions fail in that context. Phone agents need server-side tools for the actions they must perform.
My practical recommendation is to begin with a read-only lookup tool, then add a narrowly scoped write action with explicit confirmation. Give booking or account-changing tools authorization checks, request identifiers and a clear result contract. Spoken fluency is not evidence that a transaction succeeded.
Saved agent configurations are versioned, and sessions can pin a version. That provides a useful control for testing an updated prompt or voice without silently changing the production configuration.
Check these migration limits before changing model IDs
The Voxtral 2 migration guide documents concrete integration changes:
- Word timestamps are unavailable. Offline requests accept segment timestamps; requesting word timestamps produces a 400 error. Word-highlight interfaces need a different implementation strategy.
- Realtime settings lock after the first audio chunk. Send configuration before streaming audio rather than relying on later session updates.
- Speaker segments may overlap. A transcript interface should preserve overlapping speech instead of assuming strictly sequential speakers.
- Diarization and turn detection cannot share one realtime session. Both capabilities exist, but enabling both together returns an error.
Also distinguish transcription from answering questions about audio. At launch, the documented audio-chat alias voxtral-small-latest still refers to Voxtral 2. A familiar alias does not establish that every audio workflow has moved to the new generation.
Speech controls: use the documented launch features
The text-to-speech overview documents preset or cloned voices, inline emotion and pace tags, and pronunciation controls. It explicitly marks natural-language style instructions as unavailable at launch, despite broader wording in the announcement.
Use the feature matrix and actual API behavior as the integration contract. For cloned voices, obtain the speaker’s explicit permission and record its scope; access to a recording is not permission to imitate a person.
Hosting and production readiness need separate checks
The announcement lists European, US and global API endpoints. Do not infer that every request automatically stays in Europe or Canada. The documented SDK pipeline runs models at Mistral; installing a local SDK is not local model inference.
An on-premises or private-cloud route is offered through sales. That should not be described as unrestricted open weights or an automatically available self-hosted download.
For a pilot, I would check endpoint selection, retention settings, transcript access, consent, redaction, interruption handling, tool authorization, human handoff and total latency. Include noisy audio, names, numbers, silence and reconnects. Mistral’s latency claims are not an end-to-end guarantee for your network and tools.
For implementation support, see my AI consulting services and Sociaxo workflow project. The broader lesson matches my Timestral coverage: choose a model for its actual task and deployment contract, not the family name alone.
Frequently asked questions
What is Mistral Voxtral 3?
It is a transcription and speech-generation family announced with a LiveKit-based SDK for connecting speech recognition, reasoning, tools and spoken responses.
Can anyone connect a Voxtral 3 agent to a phone number?
SIP telephony is enterprise-only at launch through an account team. Browser sessions use a server-created connection, and phone agents need server-side tools.
Does Voxtral 3 support word timestamps?
No. The current offline migration contract supports segment timestamps; word timestamps return a 400 error. Realtime diarization and turn detection must also use separate sessions.
Does installing the SDK mean the models run locally?
No. The documented SDK uses models running at Mistral. Enterprise self-deployment is a separate sales-assisted route, not evidence of freely downloadable weights.
Frequently asked questions
What is Mistral Voxtral 3?
Voxtral 3 is a family of transcription and speech-generation models announced with a LiveKit-based voice-agent SDK. The SDK connects speech recognition, a reasoning model, tools and spoken responses.
Can anyone connect a Voxtral 3 agent to a phone number?
Mistral documents SIP telephony as enterprise-only at launch through an account team. Browser voice uses a server-created session; phone agents must use server-side tools rather than client-side functions.
Does Voxtral 3 support word timestamps?
The migration documentation says offline transcription accepts segment timestamps, not word timestamps. Requesting word timestamps returns a 400 error. Realtime diarization and turn detection also cannot share one session.
Does installing the SDK mean the models run locally?
No. The documented SDK pipeline runs models at Mistral while the application captures and plays audio. The announcement offers an enterprise self-deployment route through sales, not proof of freely downloadable weights.