Voxtral Transcribe 2 by Mistral Review in 2026: Download, Free, GitHub, User Experience and FAQs

What Is Voxtral Transcribe 2?

Voxtral Transcribe 2 is a speech to text model family built by Mistral AI, the Paris based startup that has been making waves in the AI world as Europe's answer to OpenAI. Released on February 4, 2026, this is not just one model but actually two models packaged together. You get Voxtral Mini Transcribe V2 for batch transcription (think processing a recorded meeting or podcast after the fact) and Voxtral Realtime for live streaming transcription (think real time captions during a video call).

The big selling point here is speed. Mistral claims the Realtime model can transcribe audio with less than 200 milliseconds of delay. That is fast enough that it feels almost instant. The batch model handles the heavier lifting with features like speaker diarization (figuring out who said what) and word level timestamps. Both models work across 13 languages, which covers most major global languages including English, Spanish, Chinese, Hindi, Arabic, French, German, Japanese, Korean, and more.

Icon Polls spent several weeks testing both models across different use cases before writing this review. Here is what we found.

How to Download Voxtral Transcribe 2

One of the things Mistral got right with this release is accessibility. The Voxtral Realtime model is completely open weight under the Apache 2.0 license. That means you can download it for free and run it on your own hardware without paying Mistral anything. The model weights are available on Hugging Face under the name Voxtral-Mini-4B-Realtime-2602 and the download comes in at about 8.87 GB.

For the batch model (Voxtral Mini Transcribe V2), things are a bit different. That one is only available through Mistral's API. You cannot download it and self host it. You will need a Mistral API key, and you will be paying $0.003 per minute of audio processed, which works out to roughly $0.18 per hour. That is cheap compared to most competitors, but it is not free.

If you just want to try it out without any setup, Mistral has an audio playground in their Mistral Studio platform where you can upload audio files and get transcripts back with diarization and timestamps. No coding required.

Is Voxtral Transcribe 2 Free?

This is one of the most searched questions about Voxtral Transcribe 2, so let us be clear about it. The Voxtral Realtime model is free to download and use. It is open weight under Apache 2.0, which means you can use it in personal projects, commercial products, or anything else without paying licensing fees. The catch is that you need your own hardware to run it. You will need a GPU with enough memory to load a 4 billion parameter model, which typically means at least 16 GB of VRAM.

The batch model (Mini Transcribe V2) is not free. It runs through Mistral's API at $0.003 per minute for batch processing and $0.006 per minute for realtime API access. There is no free tier mentioned in Mistral's documentation, though you can test it in their audio playground.

So the short answer is: partly free, partly paid. It depends on which model you want and whether you have the hardware to run the open weight version yourself.

Voxtral TTS: The Text to Speech Side

A lot of people searching for Voxtral Transcribe 2 also want to know about Voxtral TTS, so it is worth covering here. About six weeks after releasing Transcribe 2, Mistral dropped Voxtral TTS on March 23, 2026. This is a separate product that does the opposite job. Instead of turning speech into text, it turns text into speech.

Voxtral TTS is a 4 billion parameter model that generates natural sounding speech in 9 languages. It supports 20 preset voices and has a zero shot voice cloning feature where you can feed it just 3 seconds of someone's voice and it will generate new speech that sounds like that person. The API pricing is $16 per million characters and it is also available as open weights on Hugging Face under a CC BY-NC license.

Together, Voxtral Transcribe 2 and Voxtral TTS give Mistral a complete audio stack. You can transcribe speech to text and then generate speech from text, all within one ecosystem. For developers building voice agents or customer service bots, this is a pretty compelling package.

Voxtral Transcribe 2 on GitHub

Mistral itself does not maintain a dedicated GitHub repository for Voxtral Transcribe 2. However, the open source community has been active. There are several community built projects on GitHub that use the Voxtral Realtime model, including local transcription apps for macOS, Rust implementations that run in the browser, and integration examples with frameworks like vLLM and ExecuTorch.

One notable project is a complete local transcription solution that combines a macOS app with a GPU accelerated server powered by the Voxtral model. It runs entirely on your own hardware with zero cloud dependencies, which is great for privacy conscious users.

For developers looking to integrate Voxtral into their own projects, the Hugging Face model card provides code examples using Python and the Transformers library (version 5.2.0 and above). There is also a C implementation called voxtral.c for those who want maximum performance on embedded devices.

Voxtral Transcribe 2 on Hugging Face

Hugging Face is where you will find the official open weight model for Voxtral Transcribe 2. The model is listed under mistralai/Voxtral-Mini-4B-Realtime-2602 and it is released in BF16 format under the Apache 2.0 license. The model page includes code examples, deployment instructions for vLLM, and compatibility notes for ExecuTorch (for on device deployment).

The Hugging Face page also hosts the Voxtral TTS model under mistralai/Voxtral-4B-TTS-2603, along with community contributed quantized versions that reduce the model size from about 8 GB down to roughly 2.67 GB for easier deployment on smaller devices.

If you are a developer, Hugging Face is honestly the best starting point. The documentation there is solid and the community discussions on the model page can help you troubleshoot common issues.

User Experience: What It Is Actually Like to Use

After testing Voxtral Transcribe 2 across multiple scenarios, Icon Polls has a mixed but mostly positive impression. Here is the breakdown.

The accuracy is genuinely impressive for English content. In our tests with clear audio from meetings, podcasts, and interviews, the transcriptions were clean and required very little editing. The diarization feature (identifying different speakers) worked well in the batch model, correctly attributing dialogue in multi speaker conversations most of the time.

The Realtime model is fast. There is no arguing with that. The latency really does feel close to instant, which makes it viable for live captioning and real time voice agent applications. However, the Realtime model does not support speaker diarization, which is a significant limitation if you need to know who is speaking during a live call.

Where things get a bit rough is with non English languages and noisy audio. While 13 languages are officially supported, our testing found that accuracy drops noticeably with heavily accented speech and languages other than English, Spanish, and French. Context biasing (the feature that lets you feed in specific names and jargon) helps with English but is still experimental for other languages.

The developer experience is straightforward if you are comfortable with APIs and Python. But if you are a non technical user hoping for a simple desktop app, you will be disappointed. This is very much a developer focused product. There is no polished consumer application here.

Icon Polls rates Voxtral Transcribe 2 at 3.2 out of 5. It is a strong technical product with real advantages in speed and cost, but the limited language depth, lack of consumer facing tools, and the split between free and paid models hold it back from a higher score.

Pros and Cons

On the plus side, the Realtime model being open weight and free is a huge deal. The pricing on the API is among the cheapest in the market. The speed is excellent. Speaker diarization in batch mode works well. And the fact that you can run it locally means your audio data never has to leave your device.

On the downside, only 13 languages are supported compared to 50 or more from some competitors. The Realtime model lacks diarization. Context biasing only works reliably in English. There is no consumer friendly app or interface. And the batch model requires Mistral's API, so you cannot fully self host the entire product family.

Frequently Asked Questions About Voxtral Transcribe 2

1. Is Voxtral Transcribe 2 free to use?

The Voxtral Realtime model is free to download and self host under the Apache 2.0 license. The batch transcription model (Mini Transcribe V2) is only available through Mistral's paid API at $0.003 per minute.

2. How accurate is Voxtral Transcribe 2 compared to Whisper?

According to benchmarks, Voxtral Transcribe 2 achieves a word error rate of about 5.9% on the FLEURS multilingual benchmark, compared to Whisper large v3's 7.4%. In practical terms, it is noticeably better on clean English audio, though the gap narrows with noisy or heavily accented recordings.

3. Can I run Voxtral Transcribe 2 on my own computer?

Yes, the Realtime model can be downloaded from Hugging Face and run locally. You will need a GPU with at least 16 GB of VRAM. The batch model cannot be self hosted and requires Mistral's API.

4. Does Voxtral Transcribe 2 support speaker diarization?

The batch model (Mini Transcribe V2) supports full speaker diarization and can identify different speakers in a conversation. The Realtime model does not currently support diarization, which is one of its main limitations.

5. What languages does Voxtral Transcribe 2 support?

It supports 13 languages: English, Chinese, Hindi, Spanish, Arabic, French, Portuguese, Russian, German, Japanese, Korean, Italian, and Dutch. Accuracy is strongest in English, Spanish, and French based on Icon Polls testing.

6. What is the difference between Voxtral Transcribe 2 and Voxtral TTS?

Voxtral Transcribe 2 converts speech to text (transcription). Voxtral TTS converts text to speech (voice generation). They are separate products released about six weeks apart in early 2026. Together, they form Mistral's complete audio processing stack.

7. How does Voxtral Transcribe 2 compare to Google and OpenAI?

Mistral claims Voxtral Transcribe 2 outperforms GPT 4o mini Transcribe and Gemini 2.5 Flash on accuracy benchmarks. Its pricing at $0.003 per minute is also significantly cheaper than most competitors. However, Google and OpenAI offer broader language support and more polished platform features like sentiment analysis and topic detection.

8. Is Voxtral Transcribe 2 good for meeting transcription?

For batch processing of recorded meetings, yes. The diarization and timestamp features make it well suited for meeting notes. For live meeting transcription, the Realtime model is fast enough, but the lack of real time diarization means you will not get speaker labels during the live session. Overall, Icon Polls considers it a solid option for meeting transcription, especially at its price point.

9. Where can I download Voxtral Transcribe 2?

The open weight Realtime model is available on Hugging Face at mistralai/Voxtral-Mini-4B-Realtime-2602. The batch model is accessible through Mistral's API. You can also test both models in the audio playground on Mistral Studio without any downloads.

10. Does Voxtral Transcribe 2 work offline?

The Realtime model can work completely offline once downloaded and set up on your own hardware. The batch model requires an internet connection since it runs through Mistral's cloud API.