The start-up Mistral AI, with a valuation of 240 million euros, has put Mistral 7B online, its first open source generative language model. Not only free and without moderation, Mistral 7B also stands out for its efficiency and speed which positions it against giants such as OpenAI, Microsoft, Google, and Meta.
Growtika
French start-up Mistral AI, valued at 240 million euros, has unveiled Mistral 7B, a generative language model that is open source. Free and lacking a moderation mechanism, this language model would stand out for its speed and efficiency compared to competing proprietary solutions, such as OpenAI (GPT, ChatGPT), Microsoft (Bing), Google (BERT, LaMDA, Bard) and Meta (LLaMa), all for a minimal cost.
Mistral 7B is just the beginning of the company’s journey into the world of artificial intelligence. According to the official announcement, this language model can already be used to perform a variety of tasks, such as summarizing, structuring text, and answering questions. As said above, the open source aspect of Mistral 7B, published under the Apache 2.0 license, allows unrestricted use.
With Mistral 7B, the French start-up Mistral AI, co-founded by Guillaume Lample, Arthur Mensch, and Timothée Lacroix, marks a sensational entry into the field of artificial intelligence. Released on September 27, this company’s first Broad Multi-Tasking Language Model (LLM) is with 7 billion parameters.
Absent moderation
However, the freedom offered by Mistral 7B is not without raising security questions. Security researcher Paul Röttger pointed out the absence of moderation mechanisms in the model. Through a set of 175 compiled questions and answers, it highlighted the possibility of retrieving information on criminal activities, including instructions relating to murder, suicide or drug production.
In addition, its distribution of Mistral 7B via torrent, in the form of Magnet-Linkmakes its distribution irreversible, thus preventing the subsequent addition of moderation mechanisms.
What is Mistral 7B worth?
Mistral 7B does not just compete with the LLM LLaMa 2 13B: it outperforms it according to various benchmarks, while displaying performance comparable to that of LLaMa 34B, which nevertheless has 27 billion additional parameters.
The model seems to do well particularly in the areas of reasoning, comprehension and STEM reasoning (MMLU), with performance equivalent to an LLaMa 2 three times larger. This model, optimized on a public dataset from Hugging Face and launched under the name Mistral 7B Instruct, works in French.