A few months after launching Le Chat on smartphone, Mistral AI, the French startup that wants to make a place for itself in the field of artificial intelligence, is back in the spotlight. The company has successively launched a new language model, a formula for businesses, and a chatbot specialized in writing computer code.
Mistral isn't stopping there and has also announced its first reasoning model. Called Magistral, the model is designed to answer complex questions. Before delivering an answer, the AI will reason and reflect like a human being. With Magistral, Mistral AI directly responds to one of the most widespread reasoning models in the AI market, OpenAI's o1.
Imitating Human Thought
Magistral is not designed to assist you with everyday tasks. Artificial intelligence is primarily intended for financial projections, legal research, creative writing, or computer code. It is reserved for the most complicated requests.
Magistral is designed to imitate logical and organized human thought, following clear and well-structured steps. Mistral specifies that the AI's chain of thought, which forms the basis of its reasoning, is visible to the interlocutors. It is possible to follow the AI's reasoning and understand, step by step, how it constructed the response. Like OpenAI, Mistral lets users understand how AI "thinks."
Among the strengths of the reasoning model are its ability to reason in multiple languages (English, French, Spanish, German, Italian, Arabic, Russian, and Simplified Chinese), and to demonstrate mathematical logic. Typically, a reasoning model takes its time before responding. Responses don't come as quickly as with Le Chat or other chatbots. Mistral does, however, offer features like Think Mode and Flash Answers, which will provide thoughtful responses 10 times faster.
How to test Magistral?
The startup is launching two versions of Magistral: Magistral Small, a free, open-source version, and Magistral Medium, a more powerful iteration designed to help businesses. Mistral lets you test both variants. To chat with Magistral Small, you need to download the model and run it locally on your computer. It's available at HuggingFace.
If you'd rather see what Magistral Medium has to offer, head to the Chat interface. In the chatbot interface, you simply need to enable the Pure Reflection option, which can be found in the Reflection section. We briefly tested the model. The chatbot managed to solve all the problems we posed to it, most often in less than ten seconds. It provides a detailed, clear, and readable explanation of its answer and reasoning. Mistral clarifies that this is only a preliminary version of Magistral Medium.
Source: Mistral
0 Comments