🎙️ Voxtral: Mistral's Offline Transcription Model, Tested

Whisper has owned local transcription for so long that most "run speech-to-text offline" guides are really just Whisper guides with different wrappers. Mistral's Voxtral is the first open alternative we've seen that changes the shape of the problem rather than just competing on word error rate — and it quietly became infrastructure this month when LM Studio picked it for on-device dictation in Bionic.

By the numbers

• Two sizes: Voxtral 24B for production scale, Voxtral Mini 3B for local and edge
• Both Apache 2.0, both downloadable from Hugging Face
• Voxtral Mini is a drop-in replacement for Ministral; the 24B for Mistral Small 3.1
• Voxtral Realtime runs in-browser via Transformers.js + WebGPU, using Voxtral-Mini-4B-Realtime
• Browser latency: under 500ms, with audio captured locally and never transmitted
• Compatible with high-throughput serving stacks including vLLM

The thing that makes it different

Voxtral retains the text capabilities of the language model it's built on. That sounds like a footnote and isn't.

Whisper produces a transcript. Anything you want to do with that transcript — summarise it, pull action items out, answer a question about what was said — needs a second model and a second pass. Voxtral is a language model that happens to accept audio, so transcription and understanding live in one place. For a meeting-notes or interview workflow, that collapses two hops into one.

The practical consequence: you can swap Voxtral in where you were already running Ministral or Mistral Small and gain ears, rather than standing up a separate speech service alongside your text one.

Which size do you actually want

Your setup	Pick
Edge device, laptop, or 8GB GPU	Mini 3B — this is what it was built for
16GB VRAM (our 4080 tier)	Mini 3B comfortably, with room for a text model alongside
Serving many concurrent streams	24B behind vLLM
Just want dictation, zero setup	Voxtral Realtime in a browser tab
Need maximum raw accuracy, English only	Benchmark against Whisper large-v3 before switching

That last row is deliberate. We are not going to tell you Voxtral beats Whisper on accuracy, because we haven't run a controlled word-error-rate comparison on our own audio and neither, as far as we can find, has anyone with a published methodology you could check. What we can say is that it's Apache 2.0, it's offline, it's multilingual, and it does something Whisper structurally cannot.

The browser build is the easiest demo in local AI right now

Voxtral Realtime deserves its own paragraph because it removes every excuse. No install, no Python environment, no CUDA version argument. You open a page, grant microphone access, and watch live transcription appear with sub-500ms lag — running on your GPU, through WebGPU, with nothing leaving the tab.

If you have ever tried to convince a sceptical colleague that local AI is real, this is the thirty-second version of that argument. It is also a genuinely useful dictation tool on days when you don't want to think about infrastructure.

Where it fits with what you're already running

If you have a Whisper pipeline that works, this is not an emergency migration. Our Whisper local setup and the faster-whisper comparison still stand, and faster-whisper remains the throughput king for bulk transcription of an existing archive.

Reach for Voxtral when one of these is true:

• You want transcription and reasoning about the audio in one model
• You need multilingual offline transcription without a per-language stack
• You're already serving a Mistral text model and want to consolidate
• You want an Apache 2.0 licence you can point a client's legal team at

Pair it with a local TTS model like Kokoro and the whole voice loop — hear, think, speak — runs on hardware you own, which was not true a year ago.

The honest summary

Voxtral is the first credible structural alternative to Whisper for local speech, not because it transcribes better but because it doesn't stop at transcribing. The 3B Mini is the one most readers want, the browser build is the fastest way to see it work, and Apache 2.0 means nobody has to ask permission.

Benchmark it against your own audio before you rip out a working Whisper pipeline. Everything else about it is an easy yes.