Mistral open-sources Voxtral Realtime for sub-200ms voice transcription

Voxtral Realtime is a new 4B-parameter on-device speech-to-text model from Mistral AI, built from the ground up for ultra-low-latency streaming.

Released under the permissive Apache 2.0 license, the model weights are fully open and can be deployed directly on edge devices. If you are building voice agents, live captioning tools, or privacy-first applications where raw audio cannot leave the local network, you can run it entirely on your own hardware.

Streaming by design with sub-200ms latency

Instead of taking an offline model and chopping incoming audio into artificial chunks, Voxtral Realtime uses a native streaming architecture that transcribes sound as it arrives.

The transcription delay is configurable down to sub-200ms for fast voice-agent turn-taking. At a 480ms latency setting, the model stays within 1-2% word error rate of an offline model on the FLEURS benchmark. If you push the buffer to 2.4 seconds for live subtitling, it matches the accuracy of Mistral's latest batch transcription system.

The model natively handles 13 languages, including English, Chinese, Hindi, Spanish, German, Japanese, and Korean. With a 4B parameter footprint, it fits comfortably on modern edge hardware while maintaining robust transcription quality.

A batch sibling at $0.003 per minute

Alongside the open-weights Realtime model, Mistral launched Voxtral Mini Transcribe V2 for heavy-duty batch processing. This version is available via the Mistral API and Le Chat at $0.003 per minute, while the Realtime API tier costs twice as much.

Mistral claims Mini V2 achieves around a 4% word error rate on FLEURS, beating GPT-4o mini Transcribe, Gemini 2.5 Flash, and Deepgram Nova on accuracy. It also processes audio roughly three times faster than ElevenLabs Scribe v2 at one-fifth of the price.

The batch model supports single recordings up to three hours long, word-level timestamps, and speaker diarization. Note that during overlapping speech, it typically transcribes just one speaker. It also adds context biasing for up to 100 domain-specific terms or rare names, currently optimized for English with experimental support for other languages.

You can grab the Voxtral Realtime weights today directly from the Hugging Face Hub under Apache 2.0. To experiment without code, Mistral Studio now hosts an audio playground supporting up to 10 files and 1GB per upload across standard audio formats.

Why it matters

Anyone building real-time voice bots or local audio pipelines should test Voxtral Realtime immediately. Whisper has dominated open-source speech recognition for years, but its chunk-based architecture was never well suited for snappy, conversational agents. Delivering sub-200ms latency in an open 4B package solves a major headache for on-device voice workflows.

At the same time, pricing the batch API at $0.003 per minute puts intense pressure on existing cloud transcription providers. Whether you want an Apache 2.0 model on edge hardware or an dirt-cheap cloud endpoint for processing hours of meeting recordings, Mistral just raised the bar on speech price-performance.