Mistral Voxtral, the first open source audio model

When artificial intelligence systems become more capable, speech becomes fast in the default way with which we communicate with machines. Starting French AI mistake I jumped to the first open -style sound race, with the aim of challenging the dominance of the bound companies’ systems with open weight alternatives.

On Tuesday, Mistral announced the release of Voxtral, the first family of the company to aim to companies.

The company disrupts Voxtral as the first open model able to spread “really useful speech intelligence in production”.

In other words, developers will not have to choose between an open, open system that is floundering and does not really understand what is said, and works well, but it is closed, leaving developers with a higher invoice and less control over publication.

For companies, this means that Voxstral provides an alternative to affordable prices that the company claims to be “less than half of the price” similar solutions.

Mistral says Foxeral can copy up to 30 minutes of sound. Because of the LLM spine, Small Mistra 3.1It can understand up to 40 minutes, allowing users to ask questions about the audio content, create summaries, or convert audio orders into actual time procedures such as calling applications or operating functions. Voxtral is also multi -language, with the ability to copy and understand languages including English, Spanish, French, Portuguese, Indian, German, Dutch and Italian.

The company offers two types of “speech understanding models”. The first, Voxral Small, has 24B production-scale publishing parameters, and is a competitor with writer ElevenLabs, GPT-4O-MINI, and Gemini 2.5 Flash.

The second, Voxral Mini, has 3 billion local publishing parameters and edge. There is also a very fast API version, which has been stripped of a 3B model called Voxral Mini, which has been improved for copying cases only and returns with Openai Whisper’s superiority for less than half of the price.

Users can experience free Voxtral by downloading an embrace programming interface or testing models in Chatbot Le Mistral chat. Merging API into requests from $ 0.001 per minute, according to the company.

The launch comes after a month Mistral has declared legalThe first family of thinking models that work through step -by -step problems to improve reliability.

Mistral, one of the best artificial intelligence companies in Europe, is famous for its invitation to drive open source AI models. Earlier this month, Techcrunch reported that the company is holding talks to collect up to $ 1 billion in stocks Investors such as the MGX box in Abu Dhabi.