French developer Mistral AI is releasing a new set of language models designed to bring high-end AI capabilities to more people, regardless of where they are, how reliable their internet access is, or what language they speak.
The company on Tuesday announced a new large language model, called Mistral Large 3, intended for broad, general-purpose uses. Think ChatGPT or Gemini. The other models come in a range of sizes and capabilities and are built for use on devices themselves. These smaller models can run on laptops, smartphones, in cars or on robots, and can be tuned to perform specific tasks.
All of the models are open source and open weight, meaning developers who use them can see how they work and tweak them to fit their needs. “We very deeply believe this will make AI accessible to everyone, put the AI in their hand, basically,” Guillaume Lample, cofounder and chief scientist at Mistral AI, said in an interview.
Mistral AI, founded by former Google DeepMind and Meta researchers, is not as big of a name in the US as rivals like OpenAI and Anthropic, but it is better known in Europe. Along with models available for researchers and companies, it offers a chatbot called Le Chat, which is available via browser or in app stores.
AI models designed to be multilingual
Lample said the company has a goal with its new set of models to provide high-end, frontier AI capabilities that are open source and accessible. Part of that has to do with language. Most of the popular AI models in the US are built primarily to be used in English, as are benchmarking tools that compare the capabilities of models. And while those models are capable of working in other languages and translating, they may not be quite as good as the benchmarks suggest when used in non-English languages, Lample said.
Watch this: Can AI Develop a Gambling Addiction? The AI-Fueled Browser Wars, and The Future of Work With ZDNET’s Jason Hiner | Tech Today
Mistral AI wanted its new models to work better for speakers of all languages, so it increased the amount of non-English training data in proportion to English data. “I think people usually don’t push too much on the multilingual capabilities because if they do, they will also deteriorate a little bit the performance on the popular benchmarks that everybody sees,” Lample said. “So if you want to actually make your model shine on the popular benchmarks, you have to sacrifice the multilingual (performance). And conversely, if you want the model to be really good at multilingual, then you have to give up on the popular benchmarks, basically.”
A variety of sizes for a variety of uses
In addition to the general-purpose Mistral Large 3 model, with its 675 billion total parameters, there are three smaller models called Ministral 3 — 3 billion, 8 billion and 14 billion parameters — each of which comes in three varieties, for a total of nine. (A parameter is the weight or function that tells a model how to handle its input data. The bigger models are better and more capable, but they also need more computing power and work more slowly.)
The three varieties of the smaller models break down this way: one base model that can be tweaked and adjusted by the user, one fine-tuned by Mistral to perform well, and one built for reasoning spends more time iterating and processing a query to get a better answer.
Read more: AI Essentials: 29 Ways You Can Make Gen AI Work for You, According to Our Experts
The smaller models are particularly important since many AI users want something that performs one or two tasks well and efficiently versus large and costly general models, according to Lample. Developers can customize these models for those specific jobs, and a person or a company can host them on their own servers, saving the cost of running them in a data center somewhere.
Smaller models can also operate on specific devices. A tiny one could run on your smartphone, a slightly larger one on your laptop. That has benefits for privacy and security — your data never leaves your device — as well as cost and energy savings.
A small model running on the device itself does not need internet access to work, either, which is vital when you think about AI being used in things like robots and cars, where counting on reliable Wi-Fi for things to work properly is not the case.
