Mistral introduces Voxtral 3 speech models and a voice agent SDK

Mistral presents Voxtral 3 as a foundation for service desks, contact centers and voice assistants. Alongside models that turn speech into text and text into speech, it offers an SDK: a software development kit connecting those stages with reasoning and tool use. The company offers API access and directs organizations seeking deployment on their own infrastructure to its team.

The audio documentation distinguishes recorded-file transcription, live transcription and speech generation. This matters when choosing which component to integrate: the Small Transcribe model is transcription-only at launch. Every API request needs an AI Studio key. The documentation therefore describes a developer integration, with separate models for different audio jobs.

Where the conversation runs

In the documented hosted voice-agent workflow, your application captures and plays audio; transcription, the answering model and speech generation run at Mistral. Server-side tools run there too, while client-side tools execute your application’s functions. A phone call has no such client code: the guide says client-side tools fail there, so phone agents need server-side tools. Telephony is available to enterprise customers at launch through their account team.

Existing integrations may need changes

The migration guide identifies changes beyond model names. Word-level timestamps are no longer accepted and return an error; request segment timestamps instead. In realtime sessions, send all settings before the first audio chunk, because later updates return an error event. Applications displaying speaker-separated transcripts must handle overlapping segments. Mistral also recommends measuring latency again in the application, rather than treating a model switch as an unchanged user experience.

The Decoder also covered the release.