Mistral Large is available starting today and can be accessed through Mistral’s API or Azure AI, thanks to the partnership with Microsoft. The model has a context window of 32K tokens, allowing it to process large documents and recall information accurately. It also offers precise instruction-following, giving developers the flexibility to design moderation policies and call native functions.
In benchmark tests, Mistral Large performed well, with an accuracy rate of 81.2% in the massive multitask language understanding (MMLU) benchmark, placing it just behind GPT-4. The model also outperformed rivals in language-specific tests, showcasing its capabilities in multilingual text generation.
However, coding seemed to be a weak point for Mistral Large, with an accuracy rate of 45.1% in the HumanE benchmark for coding performance. Despite this, Mistral remains confident in its offering and believes it can compete with larger models like Gemini 1.5.
In addition to Mistral Large, the company has also launched an optimized version of its smaller model, Mistral Small. This model offers improved latency and cost efficiency, serving as a middle ground between Mistral’s open-weight offering and Mistral Large.
Mistral’s strategic partnership with Microsoft is a significant milestone for the company. All of Mistral’s open and commercial models, including Mistral Large, will be available on Azure AI Studio and Azure Machine Learning. This makes Mistral only the second company to make its commercial language models accessible on Azure. Mistral’s models can be seamlessly used by Azure users with their existing credits, and the company will provide direct access to its support team for customers coming via Azure.
Mistral is also expanding its distribution channels, with plans to make its open models available on Amazon Bedrock, Amazon Web Services’ managed service for AI offerings and application development.
To showcase the capabilities of their models and attract potential customers, Mistral has launched a chat app. This multilingual conversational assistant allows users to interact with Mistral’s models in a pedagogical and fun way. However, the company cautions that the app may deliver inaccurate or outdated information in some cases, as it does not have access to the internet. Mistral is also developing an enterprise-centric version of the assistant with self-deployment capacities and fine-grained moderation.
Since its inception, Mistral has raised over $500 million in funding from notable investors like Lightspeed Venture Partners and Andreessen Horowitz. With its innovative AI models and strategic partnerships, Mistral aims to make generative AI accessible to developers and businesses worldwide.
