Voxtral 3, Voice Agents, and New Specialized Models
On this page
Mistral’s autumn updates aren’t just about Large 4. On October 20, 2026, the company renewed its entire audio family and launched a managed voice pipeline; at the same time, several specialized models and third-party models arrived in the catalog. This lesson covers everything, distinguishing what’s generally available from what remains in preview.
Voxtral 3: Four Audio Models
The Voxtral 3 family includes four models, all in public preview as of November 2, 2026.
| Model | ID | Use Case | Price |
|---|---|---|---|
| Voxtral Mini Transcribe 3 | voxtral-mini-transcribe-3-0 | File transcription, with diarization and timestamping | $0.0035 per minute |
| Voxtral Small Transcribe 3 | voxtral-small-transcribe-3-0 | Most accurate transcription in the family, with diarization | $0.004 per minute |
| Voxtral Mini Transcribe Realtime 3 | voxtral-mini-transcribe-realtime-3-0 | Real-time transcription, adjustable latency and turn detection | $0.008 per minute |
| Voxtral TTS 3 | voxtral-mini-tts-3-0 | Voice synthesis with voice cloning, 35+ languages and 50 dialects, streaming | $13 per million characters |
These models replace the 2026 audio generation—Voxtral TTS (voxtral-mini-tts-2603), Voxtral Mini Transcribe 2 (voxtral-mini-2602), and its real-time version—deprecated on October 31 and removed on April 30, 2027. Mistral notes that the family is designed for sovereign deployment: via its API hosted in Europe or through self-hosting, giving the organization control over where audio is processed and conversations are retained.
Voice Agents: The Turnkey Voice Pipeline
On the same day, October 20, Mistral launched Voice Agents, a real-time bidirectional voice pipeline exposed via the Voice Agents SDK. The SDK handles the entire chain: mic capture, transcription, response generation by a model, synthesis, and playback. The models run on Mistral, your application captures and streams the audio, and interfaces are provided for telephony, browser, and custom audio. For a contact center or hands-free assistant, this is the difference between assembling five blocks and integrating one.
New Specialized Models
Two new families appeared in public preview. Shieldstral 1.0 is a compact multimodal moderation model that classifies the safety of text and images. Timestral 1.0 is a foundation model for time series, available in four sizes, that makes forecasts without prior training and produces probabilistic outputs. For production moderation, the generally available model remains Mistral Moderation 2 (mistral-moderation-2603), which handles 128,000 tokens of context and detects jailbreak attempts.
Third-Party Models on the Mistral Platform
A quieter but significant novelty: Mistral now hosts models from other publishers. As of November 2, 2026, the catalog includes Z.ai GLM 5.3 and GLM 5.3 Flash, in general availability, and DeepSeek V4.1 Flash, in public preview, all with a 1-million-token window. They are called using the same API as Mistral models; GLM 5.2 was deprecated on September 29 and removed on October 31 in favor of GLM 5.3.
OCR 4.1
Finally, on the documents side, OCR 4.1 (mistral-ocr-4-1) has been generally available since August 31, at the same price as OCR 4.0, which was removed on September 30. It adds confidence scores at the block level; the alias mistral-ocr-latest has pointed to it since July 16.
Official sources: Changelog · Models · Voice Agents · Pricing
Key Takeaways
- Voxtral 3 (October 20, public preview): three transcription models and one voice synthesis model, replacing 2026 audio
- Voice Agents handles the entire real-time voice pipeline via an SDK
- New specialists in preview: Shieldstral (moderation) and Timestral (time series)
- Hosted third-party models: GLM 5.3, GLM 5.3 Flash, DeepSeek V4.1 Flash; OCR 4.1 in general availability