Direct answer: The 3B Voxtral weights run on a single 16GB GPU and handle 30 minutes of audio per pass; self-hosting cuts per-minute cost to zero after hardware, but you trade away managed scaling and Whisper’s 90+ language coverage.

Which Voxtral size should I self-host?

Two options ship as open weights. The 3B model runs on a 16GB card (RTX 4080 or A10) with room to spare. The larger 24B variant needs roughly 48GB of VRAM in FP16 or about 24GB with 4-bit quantization. If your clips are under 30 minutes and you don’t need maximum accuracy, 3B is the sane default — it costs a fifth of the hardware and transcribes clean speech nearly as well.

How do I set it up step by step?

1. Install runtime: pip install vllm transformers torch.
2. Pull weights: huggingface-cli download mistralai/Voxtral-Mini-3B-2507.
3. Serve it: vllm serve mistralai/Voxtral-Mini-3B-2507 --max-model-len 32768.
4. Point your client at http://localhost:8000/v1 and send audio as you would to the hosted API.

The model context is 32k tokens. Because audio tokens are heavier than text tokens, that works out to the 30-minute ceiling you’ll hit in practice.

How do I handle audio longer than 30 minutes?

Split on silence, not on fixed timestamps. Fixed 30-minute cuts will slice a sentence or a pricing decision in half. Use something like ffmpeg’s silencedetect to find natural gaps, then patch any words that landed on a boundary by re-running a short overlap window. A 15-second overlap between chunks catches most dropped words.

Mistral Voxtral vs OpenAI Whisper: which should you pick?

Pick Voxtral when you want reasoning on top of the transcript — summarizing a call, extracting action items — in one pass. Pick Whisper when you need broad language support or you’re already tooled around its ecosystem.

What are the real downsides?

Four languages is limiting if you handle global calls. There’s no diarization, so multi-speaker meetings need pyannote added on top, which roughly doubles your pipeline complexity. And the 24B variant’s accuracy edge over 3B rarely justifies the GPU bill unless you’re doing medical or legal transcription.

Is self-hosting actually cheaper?

At $0.001 per minute on the hosted API, 10,000 minutes a month costs about $10. An A10 GPU rental runs $0.60–$1.00 an hour, or $430+ monthly if left running. Self-hosting only wins on privacy or at very high volume — not on raw cost at hobby scale.

Key Takeaways

• 3B Voxtral fits a 16GB GPU; 24B needs ~48GB in FP16.
• Chunk long audio on silence, not fixed timestamps, with a 15-second overlap.
• Voxtral adds audio Q&A; Whisper covers 90+ languages but transcribes only.
• Neither does diarization — budget for pyannote if you need speaker labels.
• Hosted API is cheaper than GPU rental below roughly 5,000 minutes a month.

Direct answer: Mistral Voxtral transcribes and understands up to 30 minutes of audio per request; you call it through the Mistral API by sending a base64 audio file to voxtral-mini-latest, and it returns text plus optional answers about the audio.

What is Mistral Voxtral, exactly?

Voxtral is Mistral’s speech model that does two jobs at once: transcription (audio to text) and audio understanding (answering questions about a recording). The 24B version beats Whisper large-v3 on several benchmarks while staying small enough to self-host. You get a hosted API or downloadable weights under Apache 2.0 for the small model.

How do I get an API key?

Sign in at Mistral’s console and create a key under the API Keys section. The free tier is rate-limited but enough for testing. Paid pricing for Voxtral starts at roughly $0.001 per minute of audio, billed on the transcript length, so a one-hour podcast costs a few cents.

How do I send my first audio file?

Install the client and run one call. The API expects either a URL or base64-encoded audio.

1. Run pip install mistralai.
2. Set your key: export MISTRAL_API_KEY=your_key.
3. Send audio to the model:

from mistralai import Mistral
import base64
c = Mistral(api_key="your_key")
with open("meeting.mp3","rb") as f:
    b64 = base64.b64encode(f.read()).decode()
r = c.chat.complete(model="voxtral-mini-latest", messages=[
  {"role":"user","content":[
    {"type":"text","text":"Transcribe this meeting."},
    {"type":"audio_url","audio_url":{"url":"data:audio/mp3;base64,"+b64}}
  ]}])
print(r.choices[0].message.content)

The response is plain text. If you want a summary instead of a transcript, change the text prompt to “Summarize the decisions made.” The model reads the audio directly.

Can I self-host it?

Yes. The 3B weights run on a single 16GB GPU, which is useful when audio can’t leave your network. You lose the managed scaling and pay in setup time instead. For most teams the hosted endpoint is cheaper until volume gets serious.

Where does it fall short?

The 30-minute cap forces you to chunk longer recordings, and chunk boundaries can split a speaker’s sentence. Overlapping speakers in a noisy room degrade accuracy noticeably, and there’s no built-in speaker diarization, so you’ll bolt on something like pyannote if you need “who said what.” Language coverage is four languages, well below Whisper’s 90-plus.

Who is Mistral Voxtral for?

It fits teams already in the Mistral ecosystem that want transcription with light reasoning attached. Skip it if you need 100 languages, live streaming, or diarization out of the box.

Key Takeaways

• Mistral Voxtral handles up to 30 minutes of audio per API call.
• Hosted pricing starts around $0.001 per minute; the 3B weights are self-hostable on a 16GB GPU.
• Supported languages: English, French, Spanish, Italian — not the 90+ Whisper covers.
• No speaker diarization; plan a third-party tool if you need it.