Summary: Mistral released Voxtral Transcribe 2 on February 4, 2026 with batch and streaming models, opening the realtime weights under Apache 2.0. The API prices batch transcription at $0.003 per minute and realtime at $0.006, while Mistral reports about 4 percent word error rate on FLEURS. The 4B download gives privacy-sensitive teams a self-hosted realtime path, though real-world noise and overlapping speakers still require local testing.
Mistral put out Voxtral Transcribe 2 on February 4, 2026, and the launch carries two models at once: Voxtral Mini Transcribe V2, built for batch transcription, and Voxtral Realtime, built for streaming. The streaming model’s weights are public under Apache 2.0 — listed on Hugging Face as Voxtral-Mini-4B-Realtime-2602, an 8.87 GB download — and its 4B parameters make it small enough to run on edge devices. Through the API, batch transcription runs $0.003 per minute and realtime runs $0.006 per minute.
Mistral’s own benchmark numbers put word error rate at roughly 4% on FLEURS. The release matters to developers in two ways: the floor under speech-to-text costs moves down another notch, and open-weights streaming models get within striking distance of batch quality for the first time. Nothing in Mistral’s speech lineup has taken a bigger step since the original Voxtral shipped in July 2025.
Batch and Streaming: One Launch, Two Models
Voxtral Mini Transcribe V2 is the offline half of the pair. Its feature set covers speaker diarization with exact start and end times (when speech overlaps, it usually still resolves to one speaker), context biasing — as many as 100 custom words or phrases that steer proper nouns and terminology (tuned for English; still experimental in other languages), word-level timestamps, and 13 languages: English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian, and Dutch. Noise handling is designed for factory floors and call centers, and one request accepts as much as three hours of audio. On the streaming side sits Voxtral Realtime, whose latency can be configured below 200ms. Mistral positions both models at meeting intelligence, voice agents, contact-center automation, live subtitling, and compliance documentation — the kinds of workloads where environment noise and first-word latency weigh as heavily as raw accuracy when the model gets picked.
What Latency Costs in Accuracy
Streaming transcription always forces the same trade: how long you are willing to wait against how many words come back wrong. Mistral’s figures say that at 2.4 seconds of delay the realtime model is on par with the batch one, and even squeezed down to 480ms the WER only climbs 1 to 2 percentage points. On comparisons, Mistral claims it outperforms GPT-4o mini Transcribe, Gemini 2.5 Flash, Assembly Universal, and Deepgram Nova on accuracy, and that it runs about 3 times faster than ElevenLabs Scribe v2 for a fifth of the price. Benchmarks from the vendor should always be re-run on your side — still, being within 1 to 2 points of batch while streaming is, on its own, a real widening of what realtime applications can credibly do.
Weights You Can Download, Devices You Can Own
The release’s weightiest move is putting the realtime model’s weights out in the open. With an Apache 2.0 license, 4B parameters, and an 8.87 GB download, “realtime transcription on my own hardware” stops being an integration project and becomes a download. The deployments Mistral calls out are the privacy-sensitive ones: on-premise or private-cloud setups that need to satisfy GDPR and HIPAA. Set this against the growing split over open-source strategy through 2026 (see our opening-year outlook): the European camp continues to treat open weights as leverage against cloud lock-in, and voice data sits near the top of the sensitivity scale.
The Price Tag and What Ships Around It
Worked out per hour, $0.003 a minute comes to $0.18. Integrations hit /v1/audio/transcriptions, calling either voxtral-mini-latest or voxtral-mini-transcribe-realtime-2602. Around the API, Mistral Studio picked up an audio playground that takes up to 10 files per upload (mp3, wav, m4a, flac, ogg, capped at 1GB each). Simon Willison tried it and reported a positive result — the in-browser demo transcribes fast speech and jargon like Django and WebAssembly almost immediately, and the transcripts it produces carry speaker labels and export to plain text, SRT, or JSON.
Three Takeaways for Builders
For anyone building on this, three points. First, when transcription makes up a big share of your variable costs, redo the math at $0.003 per minute — the most useful comparison is the invoice from whoever serves that traffic today. Second, teams shipping live captions, voice agents, or recording under compliance requirements gain an open path they can host themselves at 480ms latency; traffic no longer has to pass through a cloud API, and SRT or JSON exports with word-level timestamps drop straight into subtitling and search pipelines with no post-processing stage in between. Third, test against your own data: the published benchmarks cover FLEURS, Switchboard, CallHome, and AMI, yet factory noise and speakers talking over each other still call for trials in real conditions — the diarization limit on overlapping speech is a boundary the documentation states openly.