Voxtral 3 adds a modular voice-agent SDK, with enterprise-only telephony at launch

Mistral's Voxtral 3 release adds a LiveKit-based voice-agent SDK alongside new transcription and speech-generation models. The SDK connects speech recognition, a reasoning model, tool calls and spoken replies in a streaming pipeline, with support for callers interrupting the agent.

The architecture keeps those stages separate. Developers can change the language model or voice without rebuilding the agent around a different audio stack. Mistral advertises turns completing in under a second, but that is a vendor claim rather than an independently verified latency result.

The developer documentation reveals a consequential deployment distinction. Python, Node.js and browser integrations are documented, while SIP telephony is available to enterprise customers through their account team at launch. A phone call also has no application-side code running inside it: client-side tools fail there, so a telephone agent's actions must use server-side tools.

Saved agents are versioned configurations, and sessions can pin a version rather than automatically taking the newest one. That gives teams a way to test changes while keeping a known configuration in production.

Mistral lists its voice-agent API at $0.04 per minute with Mistral Small 4 for reasoning, or $0.08 with a larger model. Full on-premises or private-cloud deployment requires contacting the company.

The useful change is the packaged conversation pipeline; the transport and tool-execution constraints still determine which applications can use it.