Paris-based AI startup Mistral AI made waves this week with the launch of its new conversational assistant, Le Chat (French for "the cat"), along with the release of its most advanced language model yet, Mistral Large. The company also announced a partnership with Microsoft to bring its models to Azure AI, establishing Mistral as a leading alternative to models like Google's GPT-4.
Mistral Large boasts top-tier reasoning abilities and is fluent in English, French, Spanish, German, and Italian. It has also been designed for complex multilingual reasoning tasks, including text understanding, transformation, and code generation.
In a comparison with other leading language models, Mistral Large ranks second, just behind GPT-4, on the Measuring Massive Multitask Language Understanding (MMLU) benchmark. The model also demonstrates strong performance on other benchmarks such as HellaSwag, WinoGrande, Arc Challenge, TriviaQA, and TruthfulQA.
Specifically, Mistral Large achieves strong performance on tests of reasoning, knowledge, coding, math and cross-lingual understanding. It only has a 32,000 token context window compared to GPT-4's 128K, but this will be plenty enough for most people with average document lengths.
To showcase the capabilities of Mistral Large and its other models, Mistral AI has launched le Chat Mistral, a multilingual conversational assistant currently in beta. Le Chat Mistral serves as an interactive and enjoyable way to explore Mistral AI's technology, offering a pedagogical approach to understanding the company's advancements. The assistant can be powered by Mistral Large, Mistral Small, or Mistral Next, a prototype model designed for brevity and conciseness.
For businesses, Mistral AI is also introducing le Chat Enterprise, which comes with self-deployment capabilities and fine-grained moderation mechanisms. This offering aims to enhance team productivity while ensuring a safe and controlled environment for AI interactions.
The company also inked a deal with Microsoft to make its models available via Microsoft Azure. This makes Mistral Large available via Azure AI Studio and Azure Machine Learning, greatly expanding Mistral’s commercial distribution.
Mistral’s self-hosted enterprise offerings will also run on Azure infrastructure for sensitive use cases. The collaboration lends further credibility to models beyond GPT-4 just as AI capabilities race toward industry adoption.
Importantly, not all of Mistral’s offerings will remain open source going forward. In an interview with Le Monde, Mistral cofounder Arthur Mensch confirmed that Mistral Large would be exclusively available via commercial partnerships rather than open sourced. He defended the decision by citing the high costs of developing powerful new models, stating that “commercial activity” is necessary to “finance the costly research required.”
This marks a notable departure from Mistral’s origins as an open source-centric company. While with their earlier less capable models, they distributed the weights openly, Mistral now sees value in retaining exclusive access for its most capable AI.
Mensch says that the intial open source access remains vital for “creating demand,” but it is clear that Mistral has firmly planted its flag as a competitive commercial AI firm with elite capabilities amongst proprietary models. Open source purists will critique the decision, but Mistral is betting that premium intellectual property on Azure will further its leadership in AI innovation.
As Mistral AI continues to push the boundaries of AI technology, the company says it remains committed to gathering feedback and improving its offerings. Users are encouraged to sign up as beta testers for le Chat Mistral and provide their insights to help shape the future of Mistral AI's products. The company is also working on reducing latency and enhancing the capabilities of its models, including the introduction of JSON format mode and function calling for better developer interactions.
With these new developments, Mistral AI is poised to make significant strides in the AI industry, offering powerful language models and conversational assistants that can help businesses and individuals alike harness the power of advanced AI technology.
This article has been updated to add context about Mistral's shift from primarily offering open weights models to proprietary ones.