voxtral-mini-2602

Overview. Voxtral Mini Transcribe 2 is Mistral AI's offline speech-to-text model, released on 4 February 2026 in the Voxtral Transcribe 2 family (API id voxtral-mini-2602). It takes audio plus optional request parameters (diarization, context bias, language, and timestamp granularity) and returns transcription text, not general chat. Developers use it for file or clip transcription when they need timestamps, language detection, speaker diarization, and structured text rather than a streaming voice model.

Key strengths. Vendor documentation describes word-level timestamps, language detection, speaker diarization, context biasing for domain vocabulary, and structured output. Mistral reports it processes audio roughly three times faster than a comparable transcription model, so a finished file comes back faster than realtime. Coverage is multilingual across the 13 languages Mistral documents: English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian, and Dutch.

When to use it. Use this for offline transcription of recordings when you need speaker-aware or structured text and interactive turnaround is enough. It fits simple speech-to-text work, not multi-step reasoning, coding, or agent workflows. The Smart Router places it in the low general-quality tier with an interactive latency class and a simple complexity fit, because domain signal is concentrated on spoken-language communication and is weak for code, math, and agentic tasks.

Limitations. This is the offline Mini Transcribe model, not the streaming Realtime sibling, so it does not provide a live token stream. It has no tool calling, vision input, audio output, or extended thinking; transcription features are steered by request parameters rather than free-text instructions. Parameter count, context window, training cutoff, and public weights for this exact API id were not confirmed, so treat it as a hosted transcription model until those details are published.