Voxtral Small 24B
A solid pick for self-hosting and full data control.
Overview
Voxtral Small builds on the Mistral Small 3 text backbone and adds a full audio stack: transcription, translation and spoken-language understanding, plus function calling straight from voice. It handles up to 30 minutes of audio for transcription and 40 minutes for understanding, across eight languages including French and German. One real limit: system prompts aren't supported yet, and running it in bf16 needs roughly 55GB of GPU RAM. Released under Apache 2.0, it runs entirely on your own infrastructure, so audio data never leaves your servers.
Skill profile
Not disclosed
Strengths
- Open-source and self-hostable
Limitations
- API pricing not disclosed
Who is it for
- you want to control cost or self-host
- you have GDPR constraints
Ideal use cases
- Multilingual meeting transcription
- Call summarization pipelines
- Voice-driven business assistants
- Voice-triggered workflow automation
Access & availability
Key specifications
Privacy
Run it locally
| Quantization | Disk | RAM / VRAM | Typical hardware |
|---|---|---|---|
| Q4 · recommended | 13.7 GB | 17 GB | 32GB PC / 24GB Mac / RTX 4070 Ti+ |
| Q8 · balanced | 25.7 GB | 30 GB | 64GB Mac / dual 24GB GPUs |
| FP16 · max quality | 48 GB | 55 GB | 96-128GB Mac Studio / 4× GPUs |
Estimates for a moderate context. Long contexts need more RAM (KV cache).
Deploy
Copy-ready commands generated from this card. Adjust context length and GPU count to your hardware.
OpenAI-compatible server for production on NVIDIA GPUs.
pip install vllm
vllm serve mistralai/Voxtral-Small-24B-2507 \
--max-model-len 32000 \
--tensor-parallel-size 1 \
--dtype autoAdvanced data · for expertsArchitecture, modalities, detailed cost, full benchmarks▾
| Arena Elo | — |
| MMLU | — |
| GPQA | — |
| HumanEval | — |
| SWE-Bench | — |
| MATH | — |