Model

Voxtral Realtime

Family
Voxtral Transcribe 2
Context window
Open weights
Yes
Release date
2026-02-04

Benchmark notes

4B parameter streaming ASR with configurable latency down to sub-200ms; 13 languages; reported within 1-2% word error rate at 480ms delay; available via API at $0.006/min and as open weights on Hugging Face.