whatmodel.sh / stt / Voxtral Mini 3B (2507)

Voxtral Mini 3B (2507)

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Voxtral Mini 3B (2507) is an open-weights speech-to-text model from Mistral, released July 2025. Its published size is 4.68B parameters. Licence: apache-2.0.

58.4±1.3#43 of 162 speech-to-text modelsfrom 24 benchmark results across 11 sources

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 3B (2507), out of 176 with a score, then more from Mistral.

#modelscorewhereRAMreleased
50Parakeet Ultra Moondream59±3local~0.7GB2026-09
51Parakeet TDT 1.1B NVIDIA59±2local~5.8GB2024-01
52Voxtral Mini 3B (2507) Mistral58±1api $0.0010/audio min +local~12GB2025-07
53Universal-Streaming (English) AssemblyAI58±3api $0.0025/audio min2025-06
54Parakeet RNNT 1.1B NVIDIA58±2local~5.6GB2024-01
more from Mistral
32Voxtral Small 24B (2507) Mistral61±2api $0.0040/audio min +local~59GB2025-07
46Voxtral Mini Transcribe V2 Mistral59±3api $0.0030/audio min2026-02
96Voxtral Mini 4B Realtime (2602) Mistral53±2api $0.0060/audio min +local~3.2GB2026-02

see all speech-to-text models · find the ones that fit your machine

$ what it's good at

multilingual transcription▂··▃▄█▆▅▄▂·▂ #12 of 36 * · 19% of its score · 2 tests

Beats Cohere Transcribe Arabic (07-2026), Qwen3-ASR-1.7B, Voxtral Mini Transcribe V2 here, though they rank higher overall. Trails SeamlessM4T v2 Large, Omnilingual ASR LLM 1B v2, Whisper large-v3 here, though they rank lower overall.

testresultvs the others in its rundifficultycountswhy
Open ASR multilingual WER5.1% ✎▂··▂▃▂█▃█▂▁▁ beats 18/29avg 8.9%▇▇▇▇▇▇▇ 19%typical
Iberian ASR benchmark · 7-language macro WER (eu, ca, gl, pt, es + de, tr)23.3% ✎█▃▃··▅▃▃··▃▃ beats 5/10avg 38%▇ 0.9%noisy
other tests20 not in a task · 80% of its score
testresultvs the others in its rundifficultycountswhy
AfriSpeech-MultiBench · 7-region African-accent mean WER22.9% ✎▂····▂···▂▆█ beats 7/11avg 46%▇▇▇▇ 10%typical
AfriSpeech-MultiBench · Parliament (legal domain) WER21.1% ✎▂▂····▂·▅·▃█ beats 6/12avg 52%▇▇▇▇ 8.9%sharp
LibriSpeech test-clean WER +11.9% ✎▁···▁▁▂·▄▄██ beats 30/56avg 2.3%▇▇▇ 8.0%blunt · several runs
Mega-ASR adverse-condition suite · mean WER (CHiME-4, VOiCES, NOIZEUS)12.2% ✎▄····█▄▄···▄ beats 1/5avg 19%▇▇ 6.1%typical
Open ASR English short-form WER (March 2026)7.0% ✎▁▁▁▃··▁▁▂██▆ beats 59/86avg 8.2%▇▇ 5.5%blunt
SPGISpeech test WER lab2.2% ✎▄▄··▄██··▄██ beats 8/11avg 5.4%▇▇ 4.6%sharp
SwitchBoard WER lab11.3% ✎▄······▄···█ beats 2/3avg 22%▇▇ 4.2%typical
CHiME-4 WER lab10.6% ✎█··█·█·····█ beats 2/3avg 18%▇▇ 3.9%typical
12 more tests
testresultvs the others in its rundifficultycountswhy
Earnings-21 10-minute long-form WER lab9.6% ✎█···█·····██ beats 2/3avg 16%▇ 3.6%typical
FLEURS English test WER +13.9% ✎▁··▁▁▁▁▁▁▃█▃ beats 53/56avg 5.8%▇ 3.6%blunt · several runs
Earnings-22 10-minute long-form WER lab11.8% ✎█····█·█···█ beats 3/3avg 21%▇ 3.5%typical
GigaSpeech test WER lab10.7% ✎▁····▁··▁▄█▂ beats 4/17avg 19%▇ 3.5%typical
LibriSpeech test-other WER lab4.0% ✎▁▁▁▂▅▃▅▄█▂▂▁ beats 36/54avg 6.0%▇ 3.0%typical
Open ASR Leaderboard – Chilean Spanish · 3-dataset average WER8.3% ✎▂▂·····▂▂·█▂ beats 4/8avg 13%▇ 2.6%typical
Danish ASR Leaderboard · mean WER over 5 Danish test sets (RyeAI)25.5% ✎▂▂···▃▂▅█▃▂▂ beats 10/16avg 32%▇ 2.1%blunt
VoxPopuli English test WER lab6.9% ✎█▄▄··████▄██ beats 7/16avg 11%▇ 1.8%typical
Betthupferl (German dialect ASR benchmark) · Standard German audio → Standard German transcript WER12.1% ✎▄·█·▄▄▄▄·█▄▄ beats 8/10avg 15%▇ 1.6%blunt
Open Universal Arabic ASR Leaderboard · 6-test-set average WER42.6% ✎▃·▂▂▅█▅▃▂▅▂▂ beats 4/22avg 40%▇ 1.6%blunt
Betthupferl (German dialect ASR benchmark) · dialect audio → Standard German transcript WER34.0% ✎▃█····▃▅▃·▅▃ beats 8/10avg 41%▇ 0.7%blunt
Speech Accessibility Project dysarthric speech · zero-shot WER (rated subset, N=11,218)15.3% ✎▂·▂··▂····▆█ beats 5/9avg 18%▇ 0.6%blunt

$ where to run it

Through an API

api · DeepInfra · $0.0010/audio min · deepinfra.com ↗ · price sourceUSD per input audio minute: DeepInfra lists mistralai/Voxtral-Mini-3B-2507 at $0.00100 / minute (page pricing data: input_length billing; not deprecated in DeepInfra's catalogue). Mistral's own API retired voxtral-mini-2507 on 2026-05-31; DeepInfra serves the same open-weights checkpoint. Product: DeepInfra hosted inference (automatic speech recognition). API model: mistralai/Voxtral-Mini-3B-2507. Settings: standard serverless. Price basis: published list rate. Currency: USD. DeepInfra official model page, observed 2026-09-25.

Other ways to run it

version

  1. Run it as published · needs ~11.73 GB, NVIDIA only · from huggingface.co/mistralai
    vllm serve mistralai/Voxtral-Mini-3B-2507 --tokenizer_mode mistral --config_format mistral --load_format mistral

Weights and documentation: huggingface.co/mistralai ↗

$ how the score was worked out · Mistral AI — Voxtral technica… 30%, Open ASR Leaderboard 24%, AfriSpeech-MultiBench (arXiv:… 19% and 8 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

sourceresultsshare of the scorelicence
Mistral AI — Voxtral technical report · lab-reported10▇▇▇▇▇▇▇▇▇ 30%CC BY 4.0
Open ASR Leaderboard2▇▇▇▇▇▇▇ 24%Apache-2.0
AfriSpeech-MultiBench (arXiv:2511.14255v1) · African-accented English ASR benchmark suite2▇▇▇▇▇▇ 19%CC BY 4.0
transcribe.cpp model catalog · LibriSpeech test-clean WER of validated full-precision ports2▇▇▇ 10%MIT
Mega-ASR (arXiv:2605.19833v1) · adverse-condition ASR baselines (CHiME-4, VOiCES, NOIZEUS)1▇▇ 6%CC BY 4.0
Open ASR Leaderboard – Chilean Spanish (idsudd / Alonso Astroza)1▇ 3%Apache-2.0
German dialect ASR Benchmark (Betthupferl; kamaal-08/bavarian-asr-leaderboard)2▇ 2%MIT
Danish Speech-to-Text Leaderboard (RyeAI) · results dataset1▇ 2%MIT
Open Universal Arabic ASR Leaderboard (Elm Research Center)1▇ 2%Apache-2.0
Benchmarking Automatic Speech Recognition Tools for Iberian Languages (arXiv:2609.36920v1), Table 31▇ 1%CC BY 4.0
Dysarthric ASR with clinical context on the Speech Accessibility Project (arXiv:2605.02782v2)1▇ 1%CC BY 4.0
prior · pull toward the middle▇ 1%
weighted average= 58.4
$ quantization evidence · 1 measurements

How much of the full-precision result a quantized version kept, measured within one study. Rows from the same study share a table and aren't independent.

precisiontestfull → quantizedkept
FP8 W8A8 dynamic (weights symmetric static per-channel, activations symmetric dynamic per-token; llm-compressor FP8_DYNAMIC)FLEURS English transcription WER0.0389 → 0.039598.5%source ↗

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

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