French startup Mistral AI has unveiled a new family of models designed to break down language barriers through instant voice translation. Under the names Voxtral Mini Transcribe V2 and Voxtral Realtime, these tools promise a response speed that puts the European company ahead of US giants in specific transcription and translation tasks. The goal is to enable fluid communication between people who speak 13 different languages.
The Voxtral Realtime model stands out for its processing capacity in less than 200 milliseconds, a figure that allows for near real-time conversations. To put this into perspective, current solutions from competitors such as Google have a latency of approximately two seconds. According to the company’s managers, this breakthrough lays the groundwork for the problem of automatic translation between languages to be completely solved by 2026. Mistral’s local efficiency versus the cloud of large laboratories
Unlike the massive models from OpenAI or Anthropic, Mistral’s new tools have just 4 billion parameters. This reduced size is intentional and allows the AI to run directly on a laptop or smartphone. By not relying on external servers to process audio, private conversations remain within the device, ensuring privacy and allowing for use in places without coverage.
Mistral’s strategy distances itself from the race for general artificial intelligence to focus on specialist models. The company’s scientific management asserts that unlimited access to hardware can encourage less optimized training methods. Therefore, they have opted for an imaginative design and extreme refinement of training data. While other laboratories invest astronomical sums in raw power, the French firm seeks to offer the most efficient and economical alternative on the market.
This move reinforces Mistral’s position as a leader in European technological sovereignty. By offering open-source models and complying with European Union regulations, the company is positioning itself as a secure option for institutions seeking to reduce their dependence on US software. The launch demonstrates that small, regional models will play a key role in the future of the industry, offering practical and secure solutions for everyday tasks.
Translated with DeepL.com (free version)
That sounds awesome, this is the sort of application where LLMs actually have a chance to shine and benefit society
I tried them and they are not very good. Actually, I would describe them as bad.
That depends on your requirements or rather what you compare it with. Their models seem to focus on cheap agentic AI usage or even local installations. This will become relevant as soon as the AI bubble bursts and all big players have to massively adjust their prices. E.g. their latest “devstral 2 small” model obviously has to be compared to similar small model that offer a similar FOSS permissive license. But even comparing their bigger service based offering to similar “cheap” offerings from Anthropic it seems they can compete: https://byteiota.com/mistral-devstral-2-7x-cheaper-than-claude-72-swe-bench/
Typical IT problem. If you can scale it vertically, why should I make my system more efficient? The only reason why hyperscalers are this successful, despite the high price.
Well, for OpenAI, Google, etc this is premature optimization because they have enough money.
