Microsoft has announced a partnership with French AI startup Mistral. The models will also be accessible via Azure and support training with supercomputer capabilities. This also applies to the current top model from Mistral, which is said to be in the same weight class as the market leader GPT-4.
AI developers previously employed at Meta and Google DeepMind founded Mistral in Paris in April 2023. The startup already caused a stir in the fall because it raised 385 million euros in a financing round in October 2023. In December the value was put at 2 billion euros.
The startup is symbolic of the AI boom, but the models are also considered powerful and leading, especially in Europe. In one Stanford University Rankings The Mixtral 8x7B model performs significantly better than the Luminous models from the Heidelberg company Aleph Alpha.
Mistral is positioning itself as a European OpenAI competitor
Together with the Microsoft deal, Mistral new models released. Mistral Large is the startup’s largest Large Language Model (LLM) to date, and the benchmark results presented by the company are promising. In the often-cited MMLU test (Measuring Massive Multitask Language Understanding), it is still behind GPT-4, but performs better than Anthropic’s Claude 2, Google’s Gemini 1.0 Pro and Meta’s Llama-2 model.
Mistral Large offers the classic LLM functions such as text comprehension and coding. One focus is on dealing with languages. According to the official announcement, it supports English, French, Spanish, German and Italian with a “differentiated understanding of languages and cultural context“. The size for context inputs is 32,000 tokens.
Mistral describes the precise following of instructions at the system level as a plus point, so that developers can design adequate moderation guidelines. Mistral used this feature to provide moderation for the chatbot Le Chat to develop, which is now also available.
In addition to the large version, Mistral also presented a small version of the model. It also outperforms the previous top model Mixtral 8x7B, but is said to work significantly more efficiently. It is therefore the variant for smaller tasks that can be carried out more cost-effectively. Both Mistral Large and the Small variant are compared to the older models not available as open source. The models can be accessed via an API interface via the Mistral cloud La Plateforme, via Microsoft Azure and the chat service Le Chat.
Next partnership for Microsoft
For Microsoft, this is the next partnership with an AI startup that will be intensified. The group is famous for its close cooperation with OpenAI; the company’s GPT-4 language model is also the technical basis for Microsoft’s AI assistant Copilot. In the cloud, however, Microsoft’s strategy is to position itself as broadly as possible. In addition to the OpenAI models, Llama 2 from Meta is also available, as well as the older models from Mistral since November 2023.
Now the cooperation is being taken to a new level with a partnership that is scheduled to last several years. In addition to cloud business, Mistral also receives supercomputer capacity to train the AI models. There was also a direct investment of 15 million euros, a Microsoft spokesman told the news agency Reuters.
This part of the agreement in particular is already attracting the attention of the Brussels competition authorities. The EU is already investigating whether AI partnerships like those between Microsoft and OpenAI comply with European competition rules. The Mistral deal will now also be examined as part of the investigation.