All tasks / Models
Voxtral Mini 4B Realtime (2602)
Mistral's open-weights 4B streaming transcription model under Apache-2.0.
- Mistral AI
- Open weights
- Released Feb 2026
Where it ranks
- BTranscriptionModels · Local · 8.4% on Hugging Face Open ASR Leaderboard (English, short-form average WER) (secondary benchmark)
Facts
- Vendor
- Mistral AI
- Family
- Voxtral Realtime
- Released
- Feb 2026
- Weights
- Open: download and run it yourself
- License
- Apache-2.0
- Parameters
- 4.43B
- GPU memory to run locally
- ~10 GB (Derived: about 4.43B parameters is roughly 8.9 GB of bf16 weights plus runtime overhead; not an official requirement.)
- Access
- Free (open weights)
- Last verified
- 19 Oct 2026
Benchmark results
| Benchmark | Result | Entry on the leaderboard | As of | Source |
|---|---|---|---|---|
| Artificial Analysis Speech to Text: AA-WER Streaming | 5.2% | Voxtral Mini Transcribe RealtimeAA-WER Streaming index (final transcripts), provider Mistral; embedded value 0.052393 converted to percent and rounded; page shows no last-updated date. | 19 Oct 2026 | artificialanalysis.ai |
| Coval Benchmarks: Speech-to-Text WER (24h window) | 5.4% | mistral / voxtral-mini-transcribe-realtime-2602Coval STT benchmark, mean WER over the trailing 24h window (snapshot generation 2610, as_of 2026-10-19T09:38Z), n=858 clips; rank 21 of 31; read from the public API that feeds https://benchmarks.coval.ai. | 19 Oct 2026 | benchmarks-api-6wxgp27p2a-ue.a.run.app |
| Hugging Face Open ASR Leaderboard (English, short-form average WER) | 8.4% | mistralai/Voxtral-Mini-4B-Realtime-2602Average WER over the board's default datasets (English short-form, version 17-10-2026); rank 49 of 56; read from the Space's /config table data. | 17 Oct 2026 | huggingface.co |
Where you can use it
No tracked tool lists this model yet; as open weights it runs in local runtimes.