The innovative French startup Mistral AI enters the race of artificial intelligence leaders with its newest development – the Mistral Large model. This product promises to be a worthy alternative to the most powerful and popular solutions from OpenAI and Anthropic, in particular to their GPT-4 and Claude 2 models. In addition to LLM, the startup presented Le Chat – a generative chatbot that offers new opportunities in the field of chat assistants. It is currently available in beta.
Mistral Large is focused on performing complex text tasks in a multilingual format, including text analysis, code generation, and transformation. The model is distinguished by its ability to understand in detail the grammatical and cultural-contextual features of different languages. This feature gives it an advantage in the accuracy of information processing.
The context window of Mistral Large is quite large. It can hold up to 32,000 tokens (over 20,000 words in English). Thanks to this, the model is distinguished by its ability to work effectively with large texts. It can reproduce information with high accuracy and clearly follow the given instructions.
How Mistral Large will fare against other industry leaders is currently unknown, but its developers are encouragingly optimistic about the potential of this LLM. According to the results of the MMLU tests, the model showed 81.2% accuracy, second only to GPT-4, which scored 86.4%. This distinguishes it favorably from the main competitors. For comparison, Gemini Pro 1.0 (Google) received 71.8%, and Llama 2 70B (Meta AI) – only 69.9%.
Access to the new model will be provided through APIs and Azure AI, as Mistral AI has entered into a partnership agreement with Microsoft. The startup also announced the release of a slightly simplified version of LLM – Mistral Small and a specialized chat application. Their main purpose is to demonstrate to business teams the capabilities of the company's new AI products and how to implement them into their projects.