On February 4, 2026, Mistral released Voxtral Transcribe 2 — two models in one launch: Voxtral Mini Transcribe V2 for batch transcription and Voxtral Realtime for streaming. The realtime model ships with open weights under Apache 2.0 (Voxtral-Mini-4B-Realtime-2602 on Hugging Face, an 8.87 GB download), and at 4B parameters it is sized to run on edge devices. API pricing is $0.003 per minute for batch and $0.006 per minute for realtime.
On the FLEURS benchmark, Mistral reports roughly 4% word error rate. For developers, the release matters twice over: the cost floor for speech-to-text just dropped again, and for the first time an open-weights streaming model is within striking distance of batch quality. It is also the biggest upgrade to Mistral’s speech line since the original Voxtral in July 2025.
Two Models: Batch and Streaming
The batch model, Voxtral Mini Transcribe V2, targets offline workloads. It supports speaker diarization with precise start and end times (overlapping speech typically resolves to a single speaker), context biasing with up to 100 custom words or phrases to steer proper nouns and terminology (optimized for English, experimental elsewhere), word-level timestamps, and 13 languages: English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian, and Dutch. Noise robustness is aimed at factory floors and call centers, and a single request can take up to three hours of audio. Voxtral Realtime handles the streaming side, with latency configurable to under 200ms. Mistral pitches the pair at meeting intelligence, voice agents, contact-center automation, live subtitling, and compliance documentation — workloads where the model choice is decided by environment noise and first-word latency as much as raw accuracy.
The Latency-Accuracy Tradeoff
The hard question in streaming transcription has always been “how long do you wait, and how much do you get wrong.” Mistral’s numbers: at 2.4 seconds of delay, realtime accuracy matches the batch model; compressed to 480ms, WER rises by only 1 to 2 percentage points. Against competitors, Mistral claims higher accuracy than GPT-4o mini Transcribe, Gemini 2.5 Flash, Assembly Universal, and Deepgram Nova, and roughly 3x the speed of ElevenLabs Scribe v2 at one-fifth the cost. Vendor claims deserve your own benchmarks — but “streaming within 1 to 2 points of batch” by itself moves the feasibility boundary for realtime applications meaningfully.
Open Weights and Edge Deployment
Open-sourcing the realtime model is the most consequential part of the release. Apache 2.0 licensing, a 4B parameter footprint, and an 8.87 GB download turn “run realtime transcription locally” from an integration project into a download. Mistral explicitly aims at privacy-sensitive deployments: GDPR- and HIPAA-compliant on-premise or private-cloud setups. Against the widening split over open-source strategy in 2026 (see our opening-year outlook), the European camp keeps treating open weights as leverage against cloud lock-in — and voice is among the most sensitive data modalities there is.
Pricing and the Surrounding Ecosystem
At $0.003 per minute, the math works out to $0.18 per hour. The API endpoint is /v1/audio/transcriptions, with model IDs voxtral-mini-latest and voxtral-mini-transcribe-realtime-2602. Mistral Studio also gained an audio playground: upload up to 10 files (mp3, wav, m4a, flac, ogg, 1GB each). Simon Willison tested it and came away positive — the browser demo handles rapid speech and technical jargon like Django and WebAssembly almost instantly, and the playground produces diarized transcripts exportable as plain text, SRT, or JSON.
What It Means for Developers and Product Teams
Three takeaways. First, if transcription is a large variable cost in your product, rerun the numbers against $0.003 per minute — especially against your current vendor’s invoice. Second, products needing live captions, voice agents, or compliant recording now have an open, self-hostable path at 480ms latency; you no longer have to route all traffic through a cloud API, and the SRT and JSON export with word-level timestamps slots directly into subtitling and search pipelines without a post-processing step. Third, evaluate on your own data: Mistral’s benchmarks span FLEURS, Switchboard, CallHome, and AMI, but factory noise and overlapping speakers still demand real-world testing — the diarization limit on overlapping speech is a known boundary stated plainly in the documentation.
Sources
- Voxtral transcribes at the speed of sound — Mistral AI
- Voxtral transcribes at the speed of sound — Simon Willison
- Mistral drops Voxtral Transcribe 2, an open-source speech model that runs on-device — VentureBeat
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.