Mistral launches Voxtral 3, a voice agent kit built on LiveKit

Mistral has released Voxtral 3, a speech model family paired with a voice agent SDK. It ships three transcription models, one text-to-speech model and an SDK built on LiveKit. The idea is one place for listening, reasoning, tool calls and speaking. All four models are in public preview on the Mistral API.

The transcription lineup covers different jobs. Voxtral Small Transcribe 3 is the flagship for batch transcription and diarization. Voxtral Mini Transcribe 3 is a 4B model aimed at cheaper, faster batch jobs. Voxtral Mini Transcribe Realtime 3 handles streaming over WebSocket with end-of-turn detection. Voxtral Speech 3 is the text-to-speech model. All transcription models support diarization and context biasing, which lets you pass key terms or free text to improve recognition.

On pricing, Mistral lists its voice agent API at $0.04 a minute with Mistral Small 4. It also says the stack can run on your own infrastructure, and the company describes the Mini model as suited to resource-constrained environments. Deployment options include Mistral's API endpoints in the EU, US and globally, or self-hosting arranged through Mistral sales. GPU and VRAM requirements have not been disclosed.

Early performance claims are mixed. Small Transcribe 3 scores 2.1% on the Artificial Analysis word error rate benchmark, placing third there, and Mistral says it is strongest on diarization and multilingual batch transcription. The realtime model at 240 ms latency reaches 7.4% word error rate on FLEURS 35, compared with 4.2% for Gemini 3.5 Realtime.

Our take

Packaging transcription, reasoning, tools and speech into one SDK with a self-hosting path is a real pitch for teams wary of stitching vendors together. The realtime accuracy gap against Gemini is the number to watch before committing.