Voxtral 3: Mistral's New Speech Models, What They Cost, and How They Did on Real Voice Orders

TL;DR: Voxtral 3 is Mistral's speech family released on 20 October 2026: three transcription models, a text-to-speech model and a voice-agent stack. Transcription starts at $0.0035 per minute through Mistral's EU-hosted API, and Mistral offers self-deployment through its sales team, but no public weights. We ran the three transcription models on real WhatsApp voice orders from a UAE grocery-delivery service, in Gulf Arabic, English and Urdu: the two batch models each recovered 23 of 31 ordered products (74%), ahead of Whisper large-v3 and behind Gemini 3.8 Flash.

---

What is Voxtral 3?

Voxtral is Mistral's line of speech models. The first Voxtral models came out in July 2025 under Apache 2.0. Voxtral 3 is the third generation, and the first one built as a full voice stack rather than a pair of models:

According to Mistral's announcement, the transcription models cover more than 50 languages, including the 24 official EU languages and "many low-resource Asian and Indic languages." They identify who said what (diarization) and accept key terms or free text to steer the vocabulary, which helps with product names. Mistral claims under 500 ms of latency for the realtime model, and under a second for a full agent turn. The agent stack is built on LiveKit and connects to phone lines (SIP) and browsers (WebRTC).

All the quality figures in the announcement are Mistral's own, several of them scored by a Gemini model acting as judge. Mistral gives no per-language result for Arabic.

Where does Voxtral 3 run?

This is the question that decides whether you can use it, so here is what we could verify on 22 October 2026:

• Mistral's API, with European, US and global endpoints. Mistral says you control whether any audio is retained. We ran our test on the EU endpoint with zero data retention enabled on our organisation.
• Your own infrastructure, through Mistral. The announcement says Voxtral 3 can be deployed on-premises or in a private cloud, and to "talk to our team" for that.
• No public weights. Mistral's Hugging Face page lists Voxtral Mini 4B Realtime Arabic (8 October 2026, Apache 2.0) and earlier models, but no Voxtral 3. If you need a model you can download and run without a contract, Voxtral 3 is not it, yet.

How Voxtral 3 did on real Gulf voice orders

We had a ready-made test. Earlier this month we ran nine speech-to-text models on 54 real WhatsApp voice orders from a UAE grocery-delivery service, scored on 18 orders that came in as voice only: 35 notes, about 3.5 minutes of audio, 31 products actually delivered. A product counts only if its name or an unambiguous brand appears in the transcript, in any language or script. We ran Voxtral 3 on the same notes the day after its release, with no language set and no prompt.

What Voxtral 3 did well:

• It never got stuck. Whisper repeated itself in a loop on two notes in our re-run. No Voxtral 3 model did.
• It heard the hardest note. Voxtral Small Transcribe 3 was the first model in the benchmark to read a 3-second cigarette order, catching the brand "بلاتيني" (Platinum).
• It was fast. The API returned all 35 notes in 24 seconds with Small Transcribe 3 and 15 seconds with Mini Transcribe 3.

Where it fell short:

• Urdu comes out in Hindi script. Where Whisper wrote "دو ٹومٹر" (two tomatoes) in Urdu script, Voxtral 3 wrote "दो टमाटर" in Devanagari. A model matching against a product catalogue reads either. A shop assistant in Dubai may not.
• The streaming model drifts on short notes. Voxtral Mini Transcribe Realtime 3 turned "chawal" (rice) into Spanish and wrote other short notes in the wrong language. It finished two products behind the batch models.
• Some errors are the audio's, not the model's. Voxtral 3 and Whisper alike wrote "پیاس" (thirst) for "پیاز" (onion) and heard "سيارة" (car) for "سيجارة" (cigarette).

The full comparison, with all twelve models, the Urdu finding and the limits of the test, is in Arabic speech-to-text in 2026: 12 models tested on real Gulf voice orders.

What Voxtral 3 costs at real volumes

The prices are low enough that transcription is rarely the line that matters. Our whole scored set, 3.5 minutes of audio, cost about 1.2 cents with Mini Transcribe 3. At list price:

The real cost of a voice workflow is elsewhere: turning the transcript into something your systems can use (an order, a ticket, a CRM entry), checking it, and handling the cases where the audio is unusable.

Which Voxtral 3 model to pick

• Recorded audio (voice notes, calls, meetings): Voxtral Mini Transcribe 3. On our test it matched Small Transcribe 3 product for product, ran faster and costs less.
• Live conversations: Voxtral Mini Transcribe Realtime 3, but test it on your shortest clips first. That is where it lost ground.
• Audio that cannot leave your infrastructure: an open-weight model, today Voxtral Mini 4B Realtime Arabic or Whisper large-v3, unless you contract a self-deployment of Voxtral 3 with Mistral. Our sovereign AI page explains how we decide where each use case should run.

For voice notes that mix Arabic and English, see also how we transcribe mixed Arabic-English voice messages.

What we do, and what we haven't tested

What we do. We built a WhatsApp voice-to-order engine that turns customers' voice notes into structured orders, and we benchmark speech models on real client audio before choosing one. Client audio only goes to providers with zero data retention.

What we haven't tested. We have not tried Voxtral 3's text-to-speech, its voice-agent stack or a self-deployment. Our test set is small (31 products) and comes from one business in one city: a result on your own audio is worth more than ours.

Sources

• Mistral AI, Voxtral 3, 20 October 2026 (models, prices, languages, deployment)
• Mistral AI on Hugging Face (published Voxtral weights, checked 22 October 2026)
• Our benchmark: Arabic speech-to-text in 2026: 12 models tested on real Gulf voice orders