What is the context window size of Voxtral Mini Transcribe 3?

It features a 32,768 token context window shared between audio and custom vocabulary lists.

Цены и характеристики Mistral Voxtral Mini Transcribe 3

Q: What is the FLEURS word error rate for the model?

The model achieves a FLEURS multilingual word error rate of approximately 4.5 percent.

Voxtral Mini Transcribe 3: цены и бенчмарки API — GPUniq

Данные на 25 октября 2026 г.5 источников

Черновик подготовлен AI-инструментами и отредактирован человеком. Характеристики и выводы проверены редактором GPUniq.

TL;DR

Mistral prices Voxtral Mini Transcribe 3 batch transcription at $0.0035 per minute. Realtime transcription costs $0.008 per minute. Released on October 20, 2026, the model hits a FLEURS multilingual word error rate of approximately 4.5 percent. It features a 32,768 token context window and segment-level timestamps.

API Pricing

Batch transcription costs $0.0035 per minute of audio. Realtime transcription costs $0.008 per minute. The batch endpoint provides asynchronous transcription for recorded audio.

Batch Versus Realtime

Batch mode operates via an offline transcription endpoint for asynchronous tasks. Realtime mode handles live applications requiring low latency, though it incurs higher costs.

Transcription Mode	Price, $/min	Primary Use Case
Batch transcription	0.0035	Asynchronous recorded audio
Realtime transcription	0.0080	Low-latency live streams

FLEURS Benchmark Results

The model achieves a FLEURS multilingual transcription score of approximately 4.5 percent word error rate. It runs via API-only access with unpublished weights.

English Diarization Performance

English diarization hits an error score of approximately 18.6 percent cpWER, according to data from ai-tldr.dev. Microsoft AI Transcribe 2 sits at approximately 20.3 percent cpWER. Voxtral Small Transcribe 3 scores approximately 18.4 percent cpWER.

Key Feature Updates

Released on October 20, 2026, Voxtral Mini Transcribe 3 introduces a context window of 32,768 tokens. This space is shared between audio and custom vocabulary lists. Word-level timestamps were removed in favor of native segment-level timestamp support.

Методология

Facts are compiled from the vendor's own announcement and documentation, linked below, at the time of writing; where the vendor hasn't published a number it is marked as such. GPUniq prices, when the model is already available here, come from our live catalog.