whatmodel.sh / stt / Voxtral Mini 4B Realtime (2602)
Voxtral Mini 4B Realtime (2602)
↗ share
+ contribute
Voxtral Mini 4B Realtime (2602) is an open-weights speech-to-text model from Mistral, released February 2026. Its published size is 4.43B parameters. Licence: apache-2.0.
Score within the Open ASR English short-form WER (March 2026) comparison set only.
$ how it compares
The speech-to-text models ranked either side of Voxtral Mini 4B Realtime (2602), out of 176 with a score, then more from Mistral.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 94 |  GPT-4o Transcribe | 53±1 | api $0.0060/audio min | ||
| 95 |  LiteWhisper large-v3-turbo-acc | 53±4 | local | ||
| 96 | Voxtral Mini 4B Realtime (2602) | 53±2 | api $0.0060/audio min +local | ||
| 97 |  Parakeet primeline (German) | 53±4 | local | ||
| 98 |  Omnilingual ASR CTC 7B v2 | 52±3 | local | ||
| more from Mistral | |||||
| 32 | Voxtral Small 24B (2507) | 61±2 | api $0.0040/audio min +local | ||
| 46 | Voxtral Mini Transcribe V2 | 59±3 | api $0.0030/audio min | ||
| 52 | Voxtral Mini 3B (2507) | 58±1 | api $0.0010/audio min +local | ||
see all speech-to-text models · find the ones that fit your machine
$ what it's good at
long-form transcription█·····█···██ #1 of 4 * · 4.5% of its score · 1 test
Beats Voxtral Mini Transcribe V2, Kyutai STT 2.6B (en), GPT-4o Mini Transcribe here, though they rank higher overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Voxtral report · English long-form 4-set WER lab | 7.7% ✎ | ▄··█·······▄ beats 1/3 | 4.5% |
multilingual transcription▂··▃▄█▆▅▄▂·▂ #30 of 36 · 35% of its score · 4 tests
Beats GPT-4o Mini Transcribe here, though they rank higher overall. Trails Omnilingual ASR CTC 1B v2, Omnilingual ASR CTC 7B v2 here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Open ASR multilingual WER | 6.3% ✎ | ▂··▂▃▂█▃█▂▁▁ beats 9/29 | 25% |
| Voxtral report · Common Voice 12-language error rate lab | 15.2% ✎ | █··█·······█ beats 0/2 | 4.6% |
| Voxtral report · FLEURS 13-language error rate lab | 8.7% ✎ | █·█········█ beats 0/2 | 4.2% |
| Iberian ASR benchmark · 7-language macro WER (eu, ca, gl, pt, es + de, tr) | 48.5% ✎ | █▃▃··▅▃▃··▃▃ beats 3/10 | 1.2% |
other tests9 not in a task · 60% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Pipecat STT Benchmark · pooled semantic WER (1000 smart-turn samples, Sonnet 5.5 judge) | 5.1% ✎ | ▃·▃▅·▅▅·▃▅██ beats 0/16 | 18% |
| LibriSpeech test-clean WER +1 | 2.1% ✎ | ▁···▁▁▂·▄▄██ beats 25/56 | 10% |
| Open ASR English short-form WER (March 2026) | 7.7% ✎ | ▁▁▁▃··▁▁▂██▆ beats 45/86 | 7.3% |
| Dutch ASR Leaderboard · average WER (FLEURS, VoxPopuli, MLS nl) | 10.2% ✎ | ▄··▄█···█··▄ beats 2/6 | 5.3% |
| Voxtral report · English short-form 10-set WER lab | 8.5% ✎ | █··█·█·····█ beats 0/3 | 4.8% |
| FLEURS English test WER | 6.3% ✎ | ▁··▁▁▁▁▁▃▃█▁ beats 15/56 | 3.7% |
| FLEURS Mandarin test error rate | 10.4% ✎ | ▂▂··▂▄▂▂·▆██ beats 8/17 | 3.6% |
| GigaSpeech test WER | 10.1% ✎ | ▁····▁··▁▄█▂ beats 8/17 | 3.3% |
1 more test
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LibriSpeech test-other WER | 4.1% ✎ | ▁▁▁▂▅▃▅▄█▂▂▁ beats 34/54 | 3.2% |
$ where to run it
Through an API
api · Mistral AI · $0.0060/audio min · docs.mistral.ai ↗ · price sourceUSD per input audio minute. Realtime 2602; not batch Voxtral Mini Transcribe V2. Product: Audio transcription. API model: voxtral-mini-transcribe-realtime-2602. Settings: endpoint=/v1/audio/transcriptions, streaming=yes. Price basis: published rate. Currency: USD.
Other ways to run it
version
- Run it as published · needs ~3.16 GB, Mac only · from github.com/Blaizzy
uvx --from mlx-audio mlx_audio.stt.generate --model mlx-community/Voxtral-Mini-4B-Realtime-2602-4bit --audio audio.wav --output-path out
Weights and documentation: huggingface.co/mistralai ↗
$ how the score was worked out · Open ASR Leaderboard 32%, Pipecat STT Benchmark · seman… 18%, Voxtral Realtime 18% and 4 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 Open ASR Leaderboard | 2 | 32% |
 Pipecat STT Benchmark · semantic WER results table (pipecat-ai/stt-benchmark README) | 1 | 18% |
 Voxtral Realtime · lab-reported | 4 | 18% |
| transcribe.cpp model catalog · LibriSpeech test-clean WER of validated full-precision ports | 3 | 17% |
| VibeVoice-ASR-Streaming Technical Report (arXiv:2609.02812v2), Table 3 single-speaker recognition | 3 | 7% |
| Dutch ASR Leaderboard (tvosch) · committed result files | 1 | 5% |
| Benchmarking Automatic Speech Recognition Tools for Iberian Languages (arXiv:2609.36920v1), Table 3 | 1 | 1% |
| prior · pull toward the middle | 1% | |
| weighted average | = 52.7 |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json