whatmodel.sh / stt / Voxtral Small 24B (2507)
Voxtral Small 24B (2507)
↗ share
+ contribute
Voxtral Small 24B (2507) is an open-weights speech-to-text model from Mistral, released July 2025. Its published size is 24.3B 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 Small 24B (2507), out of 176 with a score, then more from Mistral.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 30 |  Scribe v1 | 61±2 | comparison only | ||
| 31 |  Chirp 3 | 61±2 | api price unknown | ||
| 32 | Voxtral Small 24B (2507) | 61±2 | api $0.0040/audio min +local | ||
| 33 |  Gemini 3.5 Transcribe Live | 61±3 | api $0.0090/audio min | ||
| 34 |  Granite Speech 4.1 2B Plus | 61±4 | local | ||
| more from Mistral | |||||
| 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 | ||
| 96 | Voxtral Mini 4B Realtime (2602) | 53±2 | api $0.0060/audio min +local | ||
see all speech-to-text models · find the ones that fit your machine
$ what it's good at
multilingual transcription▂··▃▄█▆▅▄▂·▂ #5 of 36 * · 32% of its score · 2 tests
Beats Cohere Transcribe Arabic (07-2026), Cohere Transcribe (03-2026), Nova-3 here, though they rank higher overall. Trails SeamlessM4T v2 Large here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Open ASR multilingual WER | 3.7% ✎ | ▂··▂▃▂█▃█▃▁▁ beats 27/29 | 25% |
| CoSE-E enterprise code-switching · 4 European pairs mean WER | 6.2% ✎ | █··██···█··█ beats 1/4 | 7.0% |
other tests14 not in a task · 67% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LibriSpeech test-clean WER +1 | 1.6% ✎ | ▁···▁▁▂·▄▄██ beats 45/56 | 11% |
| Open ASR English short-form WER (March 2026) | 6.6% ✎ | ▁▁▁▃··▁▁▂██▆ beats 68/86 | 7.5% |
| SPGISpeech test WER lab | 1.9% ✎ | ▄▄··▄██··▄██ beats 11/11 | 6.2% |
| SwitchBoard WER lab | 11.1% ✎ | ▄······▄···█ beats 3/3 | 5.7% |
| CHiME-4 WER lab | 9.6% ✎ | █··█·█·····█ beats 3/3 | 5.3% |
| Earnings-21 10-minute long-form WER lab | 9.6% ✎ | █···█·····██ beats 3/3 | 4.9% |
| FLEURS English test WER +1 | 3.5% ✎ | ▁··▁▁▁▁▁▁▃█▃ beats 55/56 | 4.9% |
| Earnings-22 10-minute long-form WER lab | 12.5% ✎ | █····█·█···█ beats 1/3 | 4.8% |
6 more tests
| test | result | vs the others in its run | counts |
|---|---|---|---|
| GigaSpeech test WER lab | 10.3% ✎ | ▁····▁··▁▄█▂ beats 5/17 | 4.7% |
| LibriSpeech test-other WER lab | 3.1% ✎ | ▁▁▁▂▅▃▅▄█▂▃▁ beats 51/54 | 4.1% |
| Danish ASR Leaderboard · mean WER over 5 Danish test sets (RyeAI) | 20.8% ✎ | ▂▂···▃▂▅█▃▂▂ beats 14/16 | 2.9% |
| VoxPopuli English test WER lab | 6.6% ✎ | █▄▄··████▄██ beats 8/16 | 2.4% |
| Open Universal Arabic ASR Leaderboard · 6-test-set average WER | 34.5% ✎ | ▃·▂▂▅█▅▃▂▅▂▂ beats 15/22 | 2.2% |
| Speech Accessibility Project dysarthric speech · zero-shot WER (rated subset, N=11,218) | 13.9% ✎ | ▂·▂··▂····▆█ beats 8/9 | 0.9% |
$ where to run it
Through an API
api · Mistral AI · $0.0040/audio min · docs.mistral.ai ↗ · price sourceUSD per input audio minute: Mistral lists Voxtral Small audio input at $0.004 per minute on /v1/chat/completions; transcript output is billed separately as output tokens ($0.4 / M tokens), text input $0.1 / M. Docs list Voxtral Small v25.07 as the current audio model. Product: Mistral API. API model: voxtral-small (25.07). Settings: standard (non-batch). Price basis: published audio-input component. Exclusions: output/text tokens, regional endpoint surcharge. Currency: USD. Mistral official API pricing page, observed 2026-09-25.
Other ways to run it
- Run it as published · needs ~58.82 GB, NVIDIA only · from huggingface.co/mistralai
vllm serve mistralai/Voxtral-Small-24B-2507 --tokenizer_mode mistral --config_format mistral --load_format mistral --tensor-parallel-size 2 --tool-call-parser mistral --enable-auto-tool-choice
Weights and documentation: huggingface.co/mistralai ↗
$ how the score was worked out · Mistral AI — Voxtral technica… 40%, Open ASR Leaderboard 33%, transcribe.cpp model catalog… 13% 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 |
|---|---|---|
 Mistral AI — Voxtral technical report · lab-reported | 10 | 40% |
 Open ASR Leaderboard | 2 | 33% |
 transcribe.cpp model catalog · LibriSpeech test-clean WER of validated full-precision ports | 2 | 13% |
| CoSE-E: A Benchmark for Code-switched Speech Evaluation in Enterprise Settings (arXiv:2609.35645v1), Table 2 | 1 | 7% |
| Danish Speech-to-Text Leaderboard (RyeAI) · results dataset | 1 | 3% |
| Open Universal Arabic ASR Leaderboard (Elm Research Center) | 1 | 2% |
| Dysarthric ASR with clinical context on the Speech Accessibility Project (arXiv:2605.02782v2) | 1 | 1% |
| prior · pull toward the middle | 1% | |
| weighted average | = 60.8 |
$ 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.
| precision | test | full → quantized | kept | |
|---|---|---|---|---|
| FP8 W8A8 dynamic (weights symmetric static per-channel, activations symmetric dynamic per-token; llm-compressor FP8_DYNAMIC) | FLEURS English transcription WER | 0.0345 → 0.0343 | 100.6% | source ↗ |
$ 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