What are Mistral Large and Le Chat, and why does Microsoft's deal with Mistral matter?

On February 26, 2024, the French company Mistral AI released Mistral Large, its most capable language model, which it ranks second only to GPT-4 among models available through an API, along with a smaller Mistral Small and a beta chat assistant called Le Chat. The same day, Microsoft announced a multi-year partnership that makes Mistral's models available on its Azure cloud.

What are Mistral Large and Le Chat, and why does Microsoft's deal…

Yesterday, February 26, 2024, Mistral AI, the Paris-based start-up known for its small open models, made three announcements at once. It released Mistral Large, its most advanced model. It opened a beta of Le Chat, its own chat assistant. And Microsoft announced a multi-year partnership that brings Mistral's models to its Azure cloud platform.

Until now, Mistral was mostly known among developers for Mistral 7B and Mixtral 8x7B, models anyone can download. Mistral Large is different: it is a commercial model you use through Mistral's service or Azure, not one you download freely.

What Mistral released

Here are the day's announcements in one table:

Announcement	What it is	Where to use it
Mistral Large	Mistral's flagship text model for complex reasoning, multilingual work and code	la Plateforme (Mistral's API), Azure, Le Chat
Mistral Small	A new model tuned for speed and lower cost	la Plateforme, Le Chat
Le Chat	A free beta chat assistant for trying Mistral's models	Mistral's website, after signing up as a beta user
Le Chat Enterprise	A business version with self-deployment and finer moderation controls	On request
Microsoft partnership	Azure computing for Mistral; Mistral models sold through Azure	Azure AI Studio and Azure Machine Learning

La Plateforme is Mistral's own service for developers, hosted on its infrastructure in Europe. An API, or application programming interface, is the connection that lets a program send text to the model and receive its reply.

How capable is Mistral Large?

Mistral says that on MMLU, a widely used test of general knowledge and reasoning across dozens of school and university subjects, Mistral Large ranks second among models available through an API, behind only OpenAI's GPT-4. Its comparison chart places it ahead of Anthropic's Claude 2, Google's Gemini Pro 1.0, OpenAI's GPT-3.5 and Meta's Llama 2 70B on that test. These are Mistral's own figures, so independent testing will matter.

The features that stand out:

Languages. Mistral says the model handles five languages natively, English, French, Spanish, German and Italian, including their grammar and cultural nuances. It reports clearly better scores than Llama 2 70B on tests in those four European languages.
Context window. It can consider 32,000 tokens at once, roughly a long report. A token is a small piece of text, often part of a word.
Function calling. The model can ask a developer's software to run a tool, such as looking up an order in a database, then use the result in its reply.
JSON mode. It can be forced to answer in valid JSON, a structured data format that other programs read easily.

Mistral Small offers the same function calling and JSON mode with lower latency, meaning it starts answering faster, and costs less. Mistral says it outperforms its own Mixtral 8x7B.

What Le Chat is

Le Chat, French for "the cat," is Mistral's answer to ChatGPT. Mistral presents it as a demonstration of what can be built with its models, and as an approachable, even playful way to explore them. You can choose which model answers: Mistral Large, Mistral Small or a prototype called Mistral Next, designed to give brief, concise answers.

Two details are worth knowing. Le Chat cannot search the internet, so its answers may be inaccurate or out of date. And it uses an adjustable moderation system that warns you, without stopping the conversation, when a discussion moves toward sensitive or controversial content. Mistral says it used Mistral Large's precise instruction following to set up that moderation.

The Microsoft partnership

Microsoft's Azure blog describes the partnership in three parts:

Computing power. Mistral gets access to Azure's AI supercomputing infrastructure to train and run its models.
Distribution. Mistral's premium models are sold to Azure customers in the model catalogs of Azure AI Studio and Azure Machine Learning, alongside OpenAI's models and others. Mistral Large is available first on Azure and Mistral's own platform.
Research. The two companies will explore training purpose-built models for selected customers, including European public-sector work.

Microsoft has offered Mistral 7B in its model catalog since November. Mistral describes Azure as its first distribution partner. For customers who need full control, Mistral also offers self-deployment, with access to model weights, on their own systems.

Why it matters

Mistral Large shows a European company competing near the top of the field less than a year after it was founded, and the Azure deal gives it access to Microsoft's large business customer base. For businesses, it adds another strong option, with particular strength in European languages and the choice of keeping data on Mistral's European infrastructure.

It also marks a change for Mistral. Its new flagship, along with Mistral Small, is offered only through its service and partners, while its open-weight models remain available. Developers who value open downloads will be watching how Mistral balances the two.

Try it: Ask the same question in English and in French, or Spanish, German or Italian, in Le Chat with Mistral Large selected. For example: "Explain in five sentences how a heat pump works." Compare the clarity of both answers.

How can you try Mistral Large and Le Chat?

The simplest way is Le Chat. Sign up as a beta user on Mistral's site, choose Mistral Large in the model menu and start with everyday tasks: a summary, a draft email, an explanation. Since it cannot browse the web, give it the text you want it to work with, and check facts yourself.

Developers can create an account on la Plateforme and call the new endpoints, mistral-large-2402 and mistral-small-2402, then test function calling and JSON mode on a real task. Azure customers can find Mistral Large in the Azure AI Studio model catalog and sign in with their Azure subscription. Mistral asks users of both the chat and the API for feedback while the products are new.