Mistral rolls out Voxtral 3 speech models and a voice-agent SDK

Mistral has introduced Voxtral 3, a family of speech models alongside a software development kit for building voice agents that can listen, reason, call tools and respond in a live conversation. The company pitches the package for uses such as contact centers, service desks and virtual assistants, where a system may need to understand a request and take several actions across connected services.

The offering includes Voxtral Small Transcribe 3 for multilingual transcription, with speaker identification, contextual prompts for domain-specific terms and end-of-turn detection. Mistral says the transcription model supports 50 languages; its real-time Voxtral Realtime Transcribe 3 supports more than 35 languages, with latency below 500 milliseconds, according to the company. The Mini transcription model has 4 billion parameters and is priced at $0.0035 per minute, Mistral says.

For spoken replies, Voxtral Mini TTS 3 generates speech in more than 35 languages and 50-plus dialects. Mistral says it can adapt emotion, pacing and emphasis to context, and offers voice adaptation using at least 30 minutes of audio. The company lists the speech service at $13 per million characters and says it costs less to run than several competing offerings; those are vendor comparisons, not independent price or quality findings.

The SDK is intended to bring transcription, reasoning, tool calls and speech into one streaming workflow, allowing an agent to retrieve information or interact with business systems during a phone or web conversation. Mistral’s product page illustrates a scenario in which an assistant translates a French referral, finds an appointment and books it. That example demonstrates the proposed workflow, but the page does not provide independent evidence of its reliability in actual deployments.

Mistral says Voxtral 3 is available through its EU-hosted API or for deployment on an organization’s own infrastructure, with control over audio handling and retention. It also claims its Small transcription model outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe on selected benchmarks. The published page describes benchmark methods and results, but the comparisons are presented by Mistral, so they should be read as the company’s own evaluations.