Skip to main content
Audio modelPublic preview

Voxtral Small Transcribe 3

Mistral AI

Released
October 20, 2026
Data date
October 24, 2026
Access
Mistral API, public preview

Voxtral Small Transcribe 3 is Mistral's transcription model with context biasing and speaker recognition. Its model card gives a rounded 32k context but no exact token count. The model supports timestamps and vocabulary to guide recognition of domain terms.

Specifications and access

SpecificationValue and source
API model ID
voxtral-small-transcribe-3-0Source
Release
Public preview since October 20, 2026Source
Access
Mistral API, public previewSource
Context length
32k, stated as a rounded value; no exact token count publishedSource
Input
Audio with optional context vocabularySource
Output
Transcript with timestamps and speaker segmentationSource
Release
2026-10-20Source
Source status
public previewSource
Parameters
not publishedSource
Context values by source
32k, stated as a rounded valueSource
Price per audio minute in USD
0.004Source
Cached price per audio minute in USD
0.0004Source
Omi report edition
Public data identifies edition 2026-10-18.1, 1,513 English recordings, 7.2096 audio hours and suite 2.0.0.Source
Model ID in the Omi evaluation
Internal Omi profile key voxtral-small-transcribe-3. It is not an established API model or checkpoint identifier. (The page does not identify this as the exact Mistral API model ID and exposes no request metadata.)Source
Omi reference corpus
1,513 recordings; 7.2096 hours; 63,815 reference words; 2,872 medical-term references; 442 drug-name references; 94 dosage/unit reference events.Source
Omi model row date
2026-10 accuracy; batch timing 2026-10-18. Accuracy is dated only to the month; the timing run has a separate day-level date.Source
Language support in the current documentation
The language table lists voxtral-small-transcribe-3-0 with support for 50 languages.Source
Model ID in the API examples
Official request docs show the model string voxtral-small-transcribe-3. (The model card uses voxtral-small-transcribe-3-0; the request documentation omits trailing '-0' and does not explain whether they are aliases.)Source
Artificial Analysis listed price per 1,000 minutes
4Source
Artificial Analysis data release date
2026-10-20Source
Artificial Analysis evaluation methodology
Approximately eight benchmark hours split among AgentTalk, VoxPopuli and Earnings22; datasets duration-weighted internally, then weighted 50/25/25.Source
Rows in the Danish results
26,780Source
Danish evaluation hardware
NVIDIA A100 80 GB, batch size 16Source
Danish evaluation submission date
2026-10-21Source