HomeAI models
Mistral: Voxtral Small 24B 2507
Mistral AI
•
voxtral-small-24b-2507
•
Released 
July 2025

Mistral AI
voxtral-small-24b-2507

Open-weight 24B model that takes audio directly for transcription, translation, and spoken Q&A across up to 40 minutes of audio.

Modality:
Text
Image
Audio
model ID
mistralai/voxtral-small-24b-2507
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

N/A
tok/s

Intelligence Index *

N/A
/ 100

Context Window *

32000
tokens

Input price

0.6
Anytoken

Output price

1.8
Anytoken
Voxtral Small: Audio Understanding and Transcription in One 24B Model Voxtral Small 24B 2507 is Mistral's larger Voxtral model. It is built on the Mistral Small 3.1 backbone with an audio encoder added. It accepts audio and text and returns text, covering transcription, speech translation, audio Q&A, and summarization without a separate ASR step. It is released under Apache 2.0 and sits above Voxtral Mini 3B in the family. It suits meeting, call, and podcast analysis where transcription and follow-up reasoning should happen in one request. Add Voxtral Small to your audio pipeline

Performance

Open-Weight Transcription Accuracy Without Giving Up Text Ability

Voxtral Small pairs speech recognition with a text-capable decoder, so one model handles transcription and follow-up reasoning. Artificial Analysis lists it with an AA-WER of 2.8%, the most accurate of the open-weights speech-to-text models it tracks. Mistral's technical report also says text benchmark performance is maintained relative to Mistral Small 3.1. In practice, pipelines that chain a separate ASR model and an LLM can collapse into a single call, though audio length is bounded by the 32K context.

Benchmarks

Voxtral Small Benchmarks: Independent WER and Vendor-Reported Audio Understanding

On the Artificial Analysis speech-to-text leaderboard, Voxtral Small has an AA-WER of 2.8%. That is first among 13 open-weights models and ahead of Voxtral Mini Transcribe 2 at 3.6%. This covers transcription only. The audio-understanding evidence is Mistral-reported: its paper says Voxtral Small beats GPT-4o mini Audio on three of seven speech-understanding benchmarks and is competitive with Gemini 2.5 Flash. Treat that second claim as vendor evidence and validate it on your own recordings.

Output Speed

*
N/A
tok/s

Intelligence Index

*
N/A
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
0
%

HLE *

Adherence to multi-step structured instructions.
0
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

Voxtral Small accepts text and audio and returns text only. The 32K-token context window has to hold audio, prompt, and response. Mistral states this covers about 30 minutes of audio for transcription or 40 minutes for understanding, so longer recordings need chunking. Function calling and structured outputs are listed in Mistral's model documentation. No reasoning mode is documented, and a verified maximum output limit was not found in official sources.
Verified Specifications — 
voxtral-small-24b-2507
*
Input modalities
Text
Image
Audio
output modalities
Text
Context window
32000
 tokens
Maximum output tokens
26214
Reasoning
No
Knowledge cutoff
July 2025
Pricing (standard)
0.6
 AnyTokens in
 / 
1.8
 AnyTokens out

Limitations & Trade-offs

Where voxtral-small-24b-2507 falls short

1
32K context caps audio length. Mistral states the window covers about 30 minutes of audio for transcription and 40 minutes for understanding. Longer meetings, call-center batches, and podcasts must be chunked, which can split speakers and context across requests. Teams processing multi-hour recordings should plan segmentation logic or consider a model with a larger context window.
2
Text-only output and no reasoning mode. Voxtral Small returns text; it does not generate speech or images, and no extended-thinking mode is documented. Its text ability comes from the Mistral Small 3.1 backbone. Complex multi-step reasoning or agentic coding is better served by a reasoning-focused model, with Voxtral handling the audio step.
3
Eight highlighted languages. Mistral's language list covers English, Spanish, French, Portuguese, Hindi, German, Dutch, and Italian. Its benchmark suite adds Arabic only for FLEURS. Accuracy on other languages is not documented in the sources reviewed, so teams serving long-tail languages should test on real audio before committing.
4
Experimental tooling and vendor-reported understanding results. Mistral's model card calls function calling experimental and notes that system prompts were not yet supported in the open-weight setup. Audio-understanding comparisons against GPT-4o mini and Gemini 2.5 Flash come from Mistral, not independent testing. Validate agent behavior and instruction adherence on your own recordings.

Best-Fit Workloads

Where this model earns its place

01

Meeting and Call Transcription


Voxtral Small has a dedicated transcription mode that auto-detects the source language, and its 32K window covers about 30 minutes of audio per request. Artificial Analysis lists an AA-WER of 2.8%, the best among open-weights models it tracks. It fits meeting recorders, interview archives, and support-call logging. Recordings beyond 30 minutes must be split into segments.

02

Audio Q&A and Summarization


Mistral says the model answers questions about audio and produces structured summaries without chaining a separate ASR model and LLM, for recordings up to 40 minutes. This suits podcast digests, call summaries, and compliance review. The understanding benchmarks are Mistral-reported, so validate quality on your own content before relying on it.

03

Voice-Triggered Actions


Mistral lists function calling straight from voice, where a spoken intent triggers backend functions or API calls without intermediate parsing. This fits voice assistants and hands-free operational workflows. The model card labels the support experimental, so add validation and confirmation steps before executing consequential actions.

04

Speech Translation


Mistral reports state-of-the-art speech translation for Voxtral Small on FLEURS-Translation, ahead of Gemini 2.5 Flash and GPT-4o-mini Audio on several pairs, including English-French, Spanish-English, and German-English. It fits multilingual meeting and support-call translation within the highlighted languages. These results are vendor-reported.

Pricing in anytokens via AnyAPI
Input
0.6
₳
Output
1.8
₳
Cache write
—
₳
Cache read
0.06
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access voxtral-small-24b-2507 via AnyAPI.ai

Access voxtral-small-24b-2507 through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate voxtral-small-24b-2507 without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access voxtral-small-24b-2507 and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test voxtral-small-24b-2507 against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use voxtral-small-24b-2507 from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use voxtral-small-24b-2507 for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Voxtral Small 24B 2507 has a 32K-token context window. Mistral states this lets it handle audio up to 30 minutes for transcription or 40 minutes for audio understanding. The window is shared by audio, prompt, and response, so longer recordings need to be split into segments.

Voxtral Small has an AA-WER of 2.8% on the Artificial Analysis speech-to-text leaderboard, ranking first among the 13 open-weights models listed. Voxtral Mini Transcribe 2 scores 3.6% on the same leaderboard. Lower is better, and results on your own audio, accents, and noise conditions may differ.

Yes. Mistral's model documentation lists function calling and structured outputs, and the model card describes triggering functions directly from voice. The model card labels function-calling support experimental, so validate tool-call arguments before executing actions. No reasoning mode is documented for this model.

Mistral highlights English, Spanish, French, Portuguese, Hindi, German, Dutch, and Italian, with automatic detection of the source language. Its FLEURS benchmarks also include Arabic. Performance on other languages is not documented in the reviewed official sources, so test them on your own audio first.

You can access Voxtral Small 24B 2507 through AnyAPI.ai, and customer-facing pricing appears in the AnyTokens pricing component on this page. The model is open-weight under Apache 2.0, so it can also be self-hosted with vLLM or Transformers. Hosted use is generally lower-cost than frontier closed multimodal models.

* Benchmark data source: Artificial Analysis artificialanalysis.ai