$ cat wiki/models/voxtral-mini-transcribe-realtime-3.md
Voxtral Mini Transcribe Realtime 3
Spec
| Attribute | Value |
|---|---|
| Developer | Mistral AI |
| Released | 2026-10-20 |
| Announced | 2026-10-20 |
| Context window | unknown |
| Pricing | $0.008/min |
| License | unknown |
| Availability | Mistral API; contact Mistral for private deployment |
| Specifications follow the launch announcement; context window and license terms are not established by that source. (source) |
Release Date
Announced and available through the API on 2026-10-20. Realtime transcription with diarization, context biasing and native end-of-turn prediction. (source)
Benchmarks
Mistral reports sub-500ms transcription latency. Its FLEURS comparison covers 35 languages and uses character error rate for Japanese, Korean and Chinese. This is transcription latency, not a complete voice-agent turn. (source)
Use Cases
Streaming voice agents that must detect turn endings and respond to interruptions. (source)
Compared To
The Voxtral voice pipeline uses separate transcription, reasoning and speech-generation models. Its reasoning stage can be replaced independently; family-level voice-agent claims are not this model's standalone benchmark. See Agents (LLM Agents). (source)
Sources
- Launch announcement and immutable capture; linked model documentation and model weights were not reviewed.
Conflicting Reports
The announcement's realtime prose says over 35 languages, while its product card says 40+ languages. The 35-language FLEURS evaluation is a separate scope. The coverage discrepancy is unresolved. (source)