2023-10-02 13:36:03
With this new tool called Mistral 7B, Mistral AI poses as a challenger to Meta’s open source model, Llama.
Things are coming together for Mistral AI. Less than a year following its creation and three months following raising 105 million euros (M€) to develop open source language models intended for businesses, the startup, now valued at 240 M€, presented its LLM open source of 7 billion parameters. A new tool initially planned for early 2024 which can be used to produce summaries, structure answers to questions, among other things.
For comparison, Open AI offers hundreds of billions of parameters. However, Mistral shows its muscles and poses as a challenger to the American giants and more precisely to all the open models available with up to 13 billion parameters and… Llama, Meta’s open source language model. Mistral claims that its tool « upgrade » Llama 2 13B on all benchmarks, performs better than Llama 1 34B “on numerous benchmarks” and is not that far from the performance of CodeLlama 7B.
Upcoming proprietary models
“In terms of reasoning, understanding and STEM reasoning (MMLU), Mistral 7B performs equivalently to a Lama 2 which would be more than 3 times its size. This is as much saved in memory as it is gained in throughput. “, details a press release.
Mistral 7B is released under the Apache 2.0 license and can be used without restriction. It can be deployed on “any cloud (AWS/GCP/Azure), using vLLM inference server and skypilot”, details the company. A model which, according to the company, represents “a first step”, in his roadmap.
For the future, Mistral AI wishes to develop proprietary models optimized for deployment on site/virtual private cloud and distributed in the form of white box solutions.
1696254598
#Mistral #unveils #open #source #language #model #LINFORMATICIEN #LINFO #CYBERRISKS #L1FO #Tech #LINFORMATICIEN #LINFORMATICIEN