Voxtral Transcribe 3
Mistral's new transcription models, among the most accurate and cheap.
Smaller version local
- Best for
- Low-cost accurate transcription by API
- On your PC
- Smaller version
- Works on
- API
- Made by
- Mistral AI
What’s good
- Third lowest error rate on Artificial Analysis
- Among the cheapest top models per minute
- Older Voxtral models are open and run locally
Downsides
- Version 3 weights are not downloadable, unlike earlier Voxtral
- Developer API only, no consumer transcription app
- Brand new, still labeled preview in places
At a glance.
- Best for
- Low-cost accurate transcription by API
- Price
- Paid
- For developers
- $0.0035 per minute (Mini), $0.004 per minute (Small)
- Version
- Voxtral Small Transcribe 3 / Mini Transcribe 3Released 2026-10
- Works on
- API
- Open source
- No
- Made by
- Mistral AI
- Website
- docs.mistral.ai
How it ranks.
- Artificial Analysis speech to text, AA-WER % (Small Transcribe 3)#3 · 2.12026-11-05
- Artificial Analysis speech to text, AA-WER % (Mini Transcribe 3)#4 · 2.22026-11-05
Can I run it myself?
Partly. A smaller version runs on your computer.
Use Older open Voxtral Mini 4B Realtime via vLLM or Transformers.
- You need
- NVIDIA GPU with 16 GB memory for the open 4B model
- Setup
- TechnicalCommand line or Python
- Laptop, 8 GBToo small
- PC, 16 GB+Too small
- Mac, 16 GBToo small
- Mac, 32 GBToo small
- Mac, 64 GB+Too small
- PC, 8 GB GPUToo small
- PC, 12 GB GPUToo small
- PC, 16 GB GPURuns
- PC, 24 GB GPURuns
- WorkstationRuns