What is Mistral 3, and why does an open-weight family from 3B to 675B parameters matter?

Yesterday, Tuesday, December 2, 2025, the French company Mistral AI released Mistral 3, its new generation of language models. The family has two branches. Mistral Large 3 is the most capable model Mistral has built so far. Ministral 3 is a set of small models designed to run on modest hardware, including laptops, robots and other devices that may not always be online.

Every model in the family is published under the Apache 2.0 license, a permissive open-source license that lets anyone use, change and build commercial products on top of it. That makes Mistral 3 an open-weight release. A model's weights are the billions of numbers it learned during training. When a company publishes them, you can download the model and run it on your own machines, instead of reaching it only through someone else's app or API (a connection that programs use to send requests to a service).

What Mistral released

TechCrunch counts ten models in total. Here is how they fit together:

A parameter is one of those learned numbers; more parameters usually means more knowledge and skill, but also more memory and computing power to run. All the models can read images as well as text, and Mistral says the family works natively in more than 40 languages. Large 3 can take in about 256,000 tokens at once, the small pieces of text a model reads, which is enough for a long report or a stack of contracts.

The three versions of each small model serve different people. A base model is the raw result of training, a starting point for companies that want to train it further. An instruct model has been tuned to follow requests and hold a conversation. A reasoning model works through a problem step by step before answering, which takes longer but improves accuracy on hard questions.

The models are available on Mistral AI Studio, Hugging Face, Amazon Bedrock, Azure Foundry and other services.

How Mistral Large 3 works

Most chatbots you have used run on dense models, where every parameter takes part in producing every word. Mistral Large 3 is a sparse mixture-of-experts model. Inside it are many smaller sub-networks, the "experts," and a routing mechanism that sends each token to only a few of them.

Think of a large clinic with dozens of specialists. Each patient sees the two or three doctors who fit their problem, not the whole staff. The clinic holds a great deal of expertise, yet each visit stays quick. In the same way, Large 3 stores 675 billion parameters of knowledge but uses only about 41 billion for each token, so it answers far faster than its full size suggests.

It is still big. Mistral trained it from scratch on 3,000 NVIDIA H200 graphics processors (GPUs, the chips that do most AI computing). Working with NVIDIA, Red Hat and the open-source vLLM project, it also published a compressed version that fits on a single server with eight A100 or H100 GPUs. The company last used the mixture-of-experts design in its Mixtral series.

Mistral reports that, after its final round of training, Large 3 matches the best open instruction-tuned models on general prompts. On LMArena, a public leaderboard where people vote between anonymous answers, it entered at second place among open models that don't use step-by-step reasoning, and sixth among all open models. Independent testing will add a fuller picture.

The small models, and a worked example

The Ministral 3 models aim at a different problem: getting useful AI onto ordinary hardware. Guillaume Lample, Mistral's co-founder and chief scientist, told TechCrunch that one GPU is enough for Ministral 3, so it can sit on an office server, a laptop or a machine in a place with poor connectivity.

Mistral also stresses efficiency. It says its instruct models often reach a comparable answer with far fewer tokens than similar models, and since you pay for and wait on every token, shorter output means lower costs and faster replies. For accuracy-first work, the 14B reasoning model scores 85% on AIME 2025, a difficult American math competition.

Here is the kind of job a small local model suits: confidential material that should never leave the building.

You are an assistant for a small accounting office. Read the client letter
below and return:
1. A three-sentence summary in plain language.
2. Every date and deadline mentioned, in a table.
3. Any question we must answer before Friday.
Answer in the same language as the letter.

[paste the letter here]

Because the model runs on the office's own computer, the letter is never sent to an outside company.

Why it matters

Mistral is small next to its American rivals. TechCrunch reports it has raised about $2.7 billion, at a valuation of $13.7 billion, a fraction of what OpenAI and Anthropic have raised. Its bet is that openness and efficiency can compete with sheer size.

Lample argues that most business tasks can be handled by small models once they are fine-tuned, meaning trained a little further on a company's own examples. Large closed models may score higher straight out of the box, he says, but customized small models can match them on specific jobs while costing less and running faster. He also ties the release to access: "We don't want AI to be controlled by only a couple of big labs."

For companies, models under their own control mean data stays in-house and work doesn't stop when an outside service goes down. Mistral is also building models with partners for robots, drones and cars.

Which Mistral 3 model should you try first?

For most individuals, start with Ministral 3 8B instruct. It is small enough for many recent computers with a graphics card, understands images and handles everyday writing, summarizing and translation. If your hardware is limited, try the 3B model; if you have a strong GPU and harder questions, move up to the 14B, and pick its reasoning version for math or logic.

Mistral Large 3 makes sense when you need top quality on long documents, coding or many languages, and you either have server-class hardware or reach it through a cloud service such as Mistral AI Studio or Amazon Bedrock.

Whichever you choose, test it on your own tasks: keep a short list of questions whose answers you know, and see how each size performs before you rely on it.