Voxtral Small 24B 2507
Mistral AI · released Jul 1, 2025 · mistralai/Voxtral-Small-24B-2507
- Type
- Open weightsApache License 2.0
- Params
- 24.3B
- Context
- 32K
about 24K words of context
Our take
Written Aug 3, 2026Voxtral Small is a 24.3-billion-parameter speech-to-text model from Mistral AI with a permissive Apache licence. It turns audio into written text with excellent accuracy on clean recordings, though its performance drops sharply in noisy or accented conditions.
Pick this for high-quality transcription of clean, prepared speech at low cost, or for batch processing where real-time speed is not required. Use it when you need open weights under a permissive licence for commercial deployment. Skip it if you are transcribing meetings, accented speech, or podcasts and video, where its error rate rises many times over.
The case for it
- Excellent accuracy on clean read-aloud audio, at 1.23% of words wrong, and still strong at 2.79% on harder read speech.
- Very fast batch transcription, at roughly 100 times real time on the benchmark we track.
- Apache 2.0 licence allows commercial use, fine-tuning and redistribution.
- Two hosted offers at identical rates, so you can switch provider without price shock.
The case against it
- Accuracy collapses in noisy conversational settings: recorded meetings see more than ten times the error rate of clean speech.
- Struggles with accented and multimedia content, with error rates roughly eight and seven times worse than on clean read speech respectively.
- Every accuracy figure we hold is English-only; nothing measures performance in other languages.
How good is it?
TranscriptionTurning speech into text3 of 5Open ASR WER · 33rd of 74
94.4%
Misses roughly one word in 18, averaged over nine English test sets.
100×59th of 62
an hour of audio in 36 seconds, on the board's own hardware. Your machine will differ.
8
Listed on the model card. The accuracy above is English only.
Percentage of words wrong on each set, lower better. Bars are scaled to this model's own worst case, not to the board.
Which languages ↓ ↑
English · French · German · Spanish · Italian · Portuguese · Dutch · Hindi
IntelligencePuzzles, maths, exam questions
Nobody we watch has scored Voxtral Small 24B 2507 for this. We would take the rating from Arena Text (overall).
CodingWriting and fixing code on its own
Nobody we watch has scored Voxtral Small 24B 2507 for this. We would take the rating from Arena Coding.
AgenticPlanning, calling tools, staying on task
Nobody we watch has scored Voxtral Small 24B 2507 for this. We would take the rating from Arena Agent (IPS).
WritingWe do not rate this
Nobody we watch has scored this model for writing. Two boards come close and neither tests writing: Arena Creative Writing asks people which of two replies they prefer, and LiveBench Language tests whether a model understood a passage.
These tests check whether a model follows instructions — a precondition for all the work above, but not a measure of how well that work is done, which is why they get no rating.
Every published score for this model9 scoresEvery figure we hold, from 9 boards, with who ran it and a link to the source — including the boards no rating above is built on.
Can you run it yourself?
- Fits in memory
- weights load entirely on the card
- Spills to system RAM
- some weights offload; much slower
- Too large
- will not load even with offload
- est
- size is calculated; the verdict could change by 10%
Comfortable fit
GeForce RTX 4090 · 24 GB
Room to spare. 5.5 GB spare means a 10% error in the size would not change the answer.
GeForce RTX 5090 · 32 GB
Room to spare. 13.5 GB spare means a 10% error in the size would not change the answer.
Apple M2 (10-core GPU) · 24 GB
Borderline fit on an estimated size. It leaves 0.7 GB spare on a size we calculated rather than measured, and a 10% error either way would change the answer.
Memory use by level
Against a 24 GB card.
All 3 sizes here are calculated, not measured. We hold no measured file for this model, so each size comes from the parameter count and every verdict above inherits that uncertainty.
Check against your own machine → · All 71 devices, with every size →
Or rent it from someone else
Cheapest of 2 live listings. Picked at the widest standard context we hold, within one quantisation slice, so the numbers beside it are a price one host actually charges.
- per 1M tokens
- $0.10 in / $0.30 out
- Context served
- 32K
- Throughput
- Not measured
| Provider | In / out per 1M tokens | Context | Throughput | Trains on prompts | Logs prompts | Zero retention |
|---|---|---|---|---|---|---|
| OpenRouter | $0.10 / $0.30 | 32K | not measured | Unknown | Unknown | Unknown |
| Mistral AI | $0.10 / $0.30 | 32K | 100 tok/s | No | Yes30 days | Unknown |
Mistral AIAcross the 2 listings we hold: 1 say they do not train on prompts, 0 say they do and 1 do not say. 0 appear in the zero-retention registry we check; the rest are unknown to us rather than confirmed either way.
What each host's API supports
From the parameter list each endpoint publishes. Streaming is omitted: nothing we hold reports it, for any model.
- ✓
- Supported
- ✗
- Not supported
- Not published
- host gave no parameter list
| Provider | Tool calling | JSON output | Strict schema |
|---|---|---|---|
| OpenRouter | ✓ | ✓ | ✓ |
| Mistral AI | ✓ | ✓ | ✓ |
Tool calling: 2 of 2 listings say yes. JSON output: 2 of 2 listings say yes. Strict schema: 2 of 2 listings say yes.
When we formed this view
Dates behind this page
Prices last checked 6h ago
What we do not know about this model yet
- We hold no measured file for it, so all 3 sizes on this page are calculated from the parameter count.
- Nothing we hold says whether an endpoint streams, so we do not show it either way.
- 1 of 2 listings do not say whether they train on prompts.
Licence and identifiers
What the licence allowsApache License 2.0, what it allows commercially, and the identifiers you need to pull this model — its Hugging Face repo, our slug and a machine-readable card.
Licence
Apache License 2.0
Fully permissive: commercial use, redistribution, and derivatives allowed. Requires attribution and a copy of the license. Includes an express patent grant.
Identifiers
- Hugging Face
- mistralai/Voxtral-Small-24B-2507
- Architecture
- Dense
- Modality record
- text+file+audio->text
- Catalogue slug
- mistralai-voxtral-small-24b-2507