What is Mistral Medium 3, and why does a mid-size model matter to businesses?

Yesterday, May 7, 2025, Mistral AI announced Mistral Medium 3, a new model aimed squarely at companies. The Paris-based startup summed up its pitch in the title of its announcement, "Medium is the new large": a model of moderate size that, by Mistral's account, performs close to much bigger and more expensive systems while costing far less to use.

Mistral, founded in 2023, has built its reputation on efficient models, starting with Mistral 7B. Medium 3 continues that line, but it is aimed less at hobbyists downloading models and more at banks, energy companies and hospitals that want AI inside their own systems.

What Mistral released

The headline numbers come from Mistral itself:

A token is a chunk of text, often part of a word; TechCrunch puts a million tokens at about 750,000 words. A GPU is the graphics chip used to run AI models; four of them is a modest server setup by the standards of large models.

The model is available from launch day through Mistral's own developer service, La Plateforme, and Amazon SageMaker, with IBM watsonx, NVIDIA NIM, Microsoft's Azure AI Foundry and Google Cloud Vertex announced as coming soon. The announcement offers no public download; companies that want to run it in their own environment are asked to contact Mistral.

The same day, according to TechCrunch, Mistral made Le Chat Enterprise generally available. It is a business version of its Le Chat assistant with a tool for building AI agents and connections to services like Gmail, Google Drive and SharePoint.

How it is meant to be used

Most companies meet AI models through an API, a web service where you send text and receive answers, paying per token. That is simple, but some organizations cannot send sensitive data, such as medical records or bank transactions, to an outside service. Others want the model to know their own vocabulary and documents in depth.

Mistral Medium 3 is designed for both cases. Besides the API, it can run in several ways:

• On-premises: on servers the company owns.
• In a VPC (virtual private cloud): a fenced-off section of a cloud provider reserved for one customer.
• Hybrid: a mix of the two.

Mistral also offers to adapt the model. Fine-tuning means continuing to train a finished model on a company's examples so it learns a style or task. Continued pretraining goes further, feeding it large amounts of a company's text so it absorbs the domain's knowledge. Mistral says the model can be connected to a company's knowledge base, its internal store of documents, so answers draw on up-to-date information.

An analogy helps. A very large model is like hiring a famous consultant: brilliant, but expensive and remote. A mid-size model that you can train and house yourself is more like a capable employee you can bring on site, teach your procedures and keep close to confidential files.

According to Mistral, beta customers in financial services, energy and healthcare are already using it to give customer service agents deeper context, to personalize business processes and to analyze complex datasets.

A worked example: what the price difference means

Price per token sounds abstract, so here is a rough calculation. Suppose a support team sends 10,000 requests a day, each with 1,500 tokens of input (the customer's message plus relevant documents) and 300 tokens of output.

That arithmetic is the heart of Mistral's argument: if a model gives nearly the quality of a premium one for a small share of the cost, many routine business tasks become cheap enough to automate.

Why it matters

Medium 3 points to a trend in the industry: the race is no longer only about the most powerful model, but also about the best result per dollar. Mistral explicitly compares its price with DeepSeek V3, a model known for low cost, and claims to beat it in both API and self-hosted use.

It also strengthens Mistral's position as a European supplier for organizations that care where their data lives. TechCrunch notes customers such as BNP Paribas, AXA and Mirakl, and reports that Mistral has raised more than 1.1 billion euros.

The usual caution applies: the benchmark results are the company's own, and independent tests will tell how well they hold up on real work. Mistral also hints that a larger model is coming in the next few weeks, and that it may be more open.

How do you decide whether a mid-size model like Mistral Medium 3 is enough for your project?

Start from the task, not the model. Collect 20 to 50 real examples of what you need, such as customer emails to answer, documents to summarize or code to review, and write down what a good answer looks like for each. Run the same examples through a mid-size model and a top-tier one, and compare the results side by side.

If the mid-size model handles most cases well, its lower cost and the option to host it yourself may outweigh the gap. If it fails on the hardest cases, consider sending only those to the larger model. Factor in privacy rules, the volume you expect and whether you will need fine-tuning, because those often matter as much as raw benchmark scores.