French AI startup Mistral unveils its largest language model to date. It is designed, among other things, to capture the diversity of European languages and cultures. As expected, it is not open source.
French AI company Mistral has unveiled its new flagship model, Mistral Large, with a 32K context window, positioning it as the EU competitor to OpenAI's GPT-4 model.
According to Mistral, the new model is suitable for complex multilingual tasks, including text comprehension, text conversion, and code generation.
It achieves good results in popular benchmarks and is the second-best model generally available via API (after GPT-4) in the renowned language understanding benchmark MMLU. Google's Ultra model is more capable but is not yet generally available via API. Mistral Large is about 20 percent less expensive than GPT-4 Turbo.
The new language model is designed to capture the cultural differences and linguistic diversity of European countries. It is fluent in English, French, Spanish, German, and Italian and, according to Mistral, has a "nuanced understanding of grammar and cultural context".
Another strength of Mistral Large, according to Mistral, is its ability to follow precise instructions that allow developers to shape their moderation policies. So-called "function calling" also makes it possible to use the language model with other tools and applications for complex interactions with internal code, APIs, or databases.
Mistral Small optimized for cost efficiency
Together with Mistral Large, the company is also introducing the optimized Mistral Small model, which is optimized for low latency and cost efficiency. In contrast to the large model, Mistral Small is open source and roughly on par with Mixtral 8x7B.
Mistral Large is now available via the in-house platform "La Plateforme" and Microsoft's Azure AI via API. According to Mistral, beta customers have already used the model successfully.
Mistral also offers the opportunity to test the model free of charge in chat mode via the newly introduced assistant. The new Small model (see below) can also be tested there.
For companies, Mistral launches "Chat Enterprise", which offers a moderation system and can be rolled out locally. The moderation system is designed to issue a warning if the LLM is in danger of generating questionable content.
With the release of Mistral Large, the French AI startup shows that it is ready to compete with big players such as OpenAI and its GPT-4 model in the field of language models - at a fraction of the budget. The focus on European languages and cultures could also make Mistral an important player in the European AI market.
However, a year after its release, GPT-4 is still unrivaled, and there aren't many arguments for using another model other than price, depending on your application, and European data protection, if you care about that.
