Mistral AI Launches Voxtral 3: Speech-to-Text, TTS, and a Voice Agent SDK
Mistral AI has released Voxtral 3, a speech model family paired with a voice agent SDK. The release ships three transcription models, one text-to-speech (TTS) model, and an SDK built on LiveKit—bringing listening, reasoning, tool calls, and speaking into a single stack. All four models are available in public preview on the Mistral API.
At a Glance
- Model size: Voxtral Mini Transcribe 3 has 4B parameters. The size of Small Transcribe 3 has not been disclosed. Both batch models list a 32k context window.
- Deployment: Available through the Mistral API (EU, US, and global endpoints) or on your own infrastructure via Mistral sales. GPU and VRAM requirements have not been disclosed. Mistral says Mini is suited to resource-constrained environments.
- Performance: Ranks in the top three on independent speech-to-text leaderboards, with its strongest results in diarization and multilingual transcription.
Why It Matters in 2026
Voice is fast becoming the default interface for agentic AI, and 2026 has seen a wave of unified speech stacks that combine transcription, reasoning, and synthesis in one pipeline. Voxtral 3 follows that pattern: instead of stitching together separate ASR, LLM, and TTS vendors, developers get a single model family and SDK designed for end-to-end voice agents—reducing latency, integration overhead, and vendor sprawl.
What's in the Release
- Three transcription models, including the 4B-parameter Voxtral Mini Transcribe 3, positioned for edge and resource-constrained deployments.
- One text-to-speech model for natural voice output.
- A voice agent SDK built on LiveKit, covering the full listen–reason–act–speak loop.
Availability
All four models are in public preview on the Mistral API, with EU, US, and global endpoints. Self-hosted deployments are available through Mistral sales; detailed hardware requirements have not yet been published.
via MarkTechPost