Mistral AI has has released a recent family of artificial intelligence models that it says will pave the way for seamless conversation between people speaking different languages.
On Wednesday, the Paris-based AI lab released two recent speech-to-text models: Voxtral Mini Transcribe V2 and Voxtral Realtime. The first one is intended for transcription of audio files in immense batches, and the second one for transcription in almost real time, within 200 milliseconds; both can translate between 13 languages. Voxtral Realtime is available for free under an open source license.
With four billion parameters, the models are diminutive enough to run locally on a phone or laptop – Mistral says it’s a first for speech-to-text – meaning private conversations don’t have to be sent to the cloud. According to Mistral, the recent models are both cheaper to operate and less error-prone than competing alternatives.
Mistral recognized Voxtral Realtime – even though the model generates text rather than speech – as a clear step towards free conversation across the language barrier, which is a problem Apple AND Google they also compete in the solution. Google’s latest model can do this translate with a two-second delay.
“We are building a system that enables seamless translation. This model basically lays the foundations for that,” says Pierre Stock, vice president of scientific operations at Mistral, in an interview with WIRED. “I think this problem will be solved in 2026.”
Founded in 2023 by Meta and Google DeepMind graduates, Mistral is one of the few European companies developing core AI models that, from a capabilities standpoint, can operate remotely close to US market leaders OpenAI, Anthropic and Google.
Without access to the same level of funding and computation, Mistral focused on improving performance through imaginative model design and careful optimization of training datasets. The goal is for micro improvements in all aspects of model development to translate into significant performance gains. “Frankly, too many GPUs make you lazy,” says Stock. “You just blindly test a lot of things, but you don’t consider what the shortest path to success is.”
Mistral’s flagship large-tongue model (LLM) does not fit competing models developed by American competitors for rugged capabilities. But the company carved out a market for itself by achieving a compromise between price and performance. “Mistral offers a more cost-effective alternative where the models are not as large but good enough and can be openly shared,” says Annabelle Gawer, director of the Center for the Digital Economy at the University of Surrey. “It may not be a Formula 1 car, but it is a very capable family car.”
Meanwhile, while its American counterparts spend hundreds of billions of dollars in the race to artificial general intelligence, Mistral is building a pipeline of specialized – if less sexy – models designed to perform narrow tasks such as converting speech to text.
“Mistral does not position itself as a niche player, but it certainly creates specialized models,” says Gawer. “As an American player with resources, you want to have a very powerful general-purpose technology. You don’t want to waste your resources on tuning it to the languages and specifics of certain sectors or geographies. Leave this type of less profitable business on the table, which will create room for mid-market players.”
As relations between the US and its European allies show signs of deterioration, Mistral is also increasingly reaching back to its European roots. “There is a trend in Europe where companies, especially governments, are taking a very close look at their dependence on US software and artificial intelligence companies,” says Dan Bieler, principal analyst at consulting firm IT PAC.
Against this background, Mistral positions itself as the safest pair of hands: a European-origin, multilingual, open-source alternative to proprietary models developed in the US. “Their question has always been: How do we build a defensible position in a market dominated by massively funded U.S. entities?” says Raphaëlle D’Ornano, founder of the consulting company D’Ornano + Co. “The approach taken by Mistral so far is that it wants to be a sovereign alternative, compliant with all regulations that may apply in the EU.”
While the performance gap with U.S. heavyweights will remain as enterprises grapple with finding a return on their AI investments and taking geopolitical context into account, smaller models tailored to industry- and region-specific requirements will have their day, Bieler predicts.
“LLM companies are giants dominating discussions, but I wouldn’t count on it always being like that,” says Bieler. “Small and more regional models will play a much bigger role in the future.”