Voxtral Small 24B Guide: Speech In, Text Out, Apache 2.0

The standard speech pipeline has two models in it: a recogniser that turns audio into text, and a language model that does something useful with the text. Every handoff between them loses information and adds latency, and you pay for and operate both.

Voxtral Small is Mistral's attempt to collapse that into one model. It takes audio directly and answers in text, so transcription, summarisation, question answering and even function calling can happen in a single request. Whether that is better than the two-stage pipeline depends on which part of the pipeline was actually hurting you.

What Voxtral Small is

Mistral released Voxtral on 15 July 2025 in two sizes: a 24B model for production use, which is this one, and a 3B Voxtral Mini aimed at local and edge deployment. Both are under Apache 2.0, a genuinely permissive licence with no user thresholds or branding clauses.

The 24B model is built on Mistral's Small 3 line (the model card credits its text abilities to Mistral Small 3.1), so it keeps that model's text skills and adds an audio input path. Its context window is 32K tokens. Mistral states it handles up to 30 minutes of audio for transcription and up to 40 minutes for understanding tasks such as summarisation or answering questions about a recording.

The Hugging Face card lists English, Spanish, French, Portuguese, Hindi, German, Dutch and Italian as supported languages, with automatic language detection. If your audio is in a language outside that list, test before you assume anything.

The dated suffix in voxtral-small-24b-2507 marks the July 2025 snapshot. Pin that identifier rather than a moving alias if you need transcripts to stay consistent over time; pinning model versions explains why.

What one model replaces

The practical gain is not mostly accuracy. It is plumbing.

• Ask questions of a recording directly. Send a meeting recording and a question in one request, rather than transcribing, chunking the transcript and prompting a second model.
• Voice to function call. Mistral highlights calling backend functions straight from spoken intent, though the model card labels function calling experimental. For a voice assistant, that removes a whole stage where a transcription error becomes a wrong tool argument.
• Structured output from audio. The OpenRouter listing shows JSON mode and structured outputs, so extracting fields from a call into a schema is one step. Structured outputs covers how to make that reliable.

What you lose is the transcript as a separate artefact. In a two-stage pipeline the transcript is something you can store, search and audit. If you need that for compliance or review, ask Voxtral for a transcript explicitly and keep it, rather than only keeping the final answer.

Accuracy claims, and how to check them

Mistral claims Voxtral outperforms Whisper large-v3 across its transcription tests and that Voxtral Small beats Whisper on every FLEURS task. It also claims transcription wins over GPT-4o mini Transcribe and Gemini 2.5 Flash, competitive audio understanding against both, and parity with ElevenLabs Scribe at less than half the price. These are vendor numbers, published largely as charts.

Speech accuracy is unusually sensitive to the audio itself: microphone quality, background noise, crosstalk, accents and domain vocabulary all move word error rate more than the gap between good models. A benchmark on clean read speech tells you little about a call centre recording.

The useful test is fifty clips from your real audio, hand-transcribed once, scored by word error rate against each candidate. Pay special attention to proper nouns, product names and numbers, which are where downstream damage happens. Vendor-claimed vs independent benchmarks covers how far to trust the published figures in the meantime.

Hosting it yourself

Apache 2.0 weights make self-hosting straightforward legally, and the hardware is the real question. The model card says running it in bf16 or fp16 needs roughly 55 GB of GPU memory, and its vLLM example spreads the model across two GPUs with tensor parallelism. That is two mid-range data-centre cards or one large one, not a laptop.

Recommended settings differ by task: temperature 0.0 for transcription, and temperature 0.2 with top-p 0.95 for chat-style audio understanding. Using the chat settings for transcription invites the model to paraphrase, which is the last thing you want in a transcript.

If 55 GB is too much, Voxtral Mini at 3B is the sibling to evaluate, and quantisation is the other lever. Which open model fits your GPU covers the memory arithmetic.

Pricing through an API

On OpenRouter, Voxtral Small is listed at $0.10 input and $0.30 output per million tokens at the time of writing. Audio is converted to tokens on the way in, so the cost of a recording depends on its length and how the endpoint meters audio; check that before projecting a monthly bill. Mistral's own announcement quoted audio API pricing starting at $0.001 per minute, the entry price across the Voxtral range rather than necessarily this model.

The 32K context is the practical ceiling. Long recordings plus a long prompt plus a long answer have to fit inside it, so for hours of audio you still need to split the input and combine results.

When to pick it

• Pick Voxtral Small when you want audio understanding, not just transcription, and would rather run one model than two. The permissive licence makes it a strong default for self-hosted voice features.
• Keep a dedicated recogniser when you need word-level timestamps, speaker diarisation or very long recordings, or when your languages fall outside the supported list.
• Look at Voxtral Mini when the job is plain transcription on constrained hardware.

For a closed alternative with audio built in, compare against GPT Audio on the same clips. Run both against your own recordings, score word error rate and task success, and let that decide.