The French startup Mistral AI has unveiled a new language model which intends to breathe new life into the small world of AI, currently dominated by big names in technology such as Microsoft and Google.
At a time when the power and the effectiveness of large linguistic models is measured by the number of parameterswhich in some way constitutes the learning data that a model uses to make its predictions, Mistral AI engineers decided to design Mistral 7B, an AI trained with fewer parameters (7 billion in this case) , but still more effective than some LLMs from OpenAI or Meta.
Mistral AI also takes LLaMA, Facebook’s large language model, as a point of comparison when presenting the capabilities of its AI. The startup’s engineers thus measured the performances of Mistral 7B and those of LLaMA 2 7B, 13B, and even LLaMA 1 34B. The Mistral 7B did not come into contact with Meta’s most efficient model, LLaMA 2 70Bbut the results obtained compared to other LLMs are truly impressive.
The AI of this French startup is more efficient than models 3 times larger
The Meta and Mistral AI models compete in categories such as reasoning, understanding, but also math, programming or the MMLU (which measures language understanding in massive multitasking mode). From these benchmarks, it appears that French Artificial Intelligence is much superior to LLaMA on all these criteria. For the same number of parameters, Mistral 7B is much better than Facebook models trained with three times more data. Thanks to these results, the company can therefore pride itself on offering “the best 7B model to date”.
Mistral 7B is just the beginning for the company. She counts design an even broader LLM in order to, why not, go and tickle ChatGPT and the 1700 billion parameters of GPT-4. The icing on the cake: if you are tempted by the experience, you can install Mistral 7B on your PC. Its designers make it available to users free of charge, under the Apache 2 license.
Source: Mistral AI