Mistral Large 4 'Le Chonk' Explained: What It Is and How to Use It
Mistral's new 1-trillion-parameter model is out as a paid preview, with free downloadable weights promised by the end of October. Here's what it does and how it compares with ChatGPT and Gemini.

In this post7 sections
On 6 October, French AI company Mistral released a public preview of Mistral Large 4, a model with about 1 trillion parameters that the company has cheerfully nicknamed "Le Chonk". The big deal: Mistral says it will publish the model's weights by the end of October, which means anyone with the right hardware can download it and run it themselves. That makes it the most capable open AI model to come out of Europe so far.
For most people it is not a ChatGPT replacement yet. Here is what it is, what it costs, how you can get your hands on it and where it really stands against ChatGPT and Gemini.

What Mistral Large 4 actually is
Mistral Large 4 is a big language model, the same kind of technology behind ChatGPT and Gemini. You give it text or images, and it replies in text. It cannot generate images or speak.
It uses a design called mixture of experts. Picture a huge team of specialists, where only the few who suit a given question get called in. The whole model has around 1 trillion parameters (the internal settings it learned during training), but only about 49 billion are switched on for each word it writes. This keeps it faster and cheaper to run than its size would suggest.
Here are the main specs from Mistral's announcement and its model documentation:
- Size: 1 trillion total parameters (the docs say 1.05 trillion), with 49 billion active per word. The docs page now says 52 billion, so Mistral's own numbers vary a little
- Input: text and images, with a 1.6 billion parameter vision part for reading images
- Output: text only
- Memory for long documents: Mistral lists a 1 million token context window, roughly a few thousand pages of text. Benchmarking firm Artificial Analysis lists 512,000 tokens, so the limit you get may depend on where you use it
- Languages: more than 160, including every official EU language
- Training: 3,800 Nvidia Grace Blackwell chips in Mistral's own European data centres, according to Mistral. VentureBeat and TechCrunch say about 4,000
Where the name 'Le Chonk' comes from
The nickname began as an online joke. As VentureBeat reports, a meme spread in June about a giant, made-up French model, and CEO Arthur Mensch joined in on X. Mistral leaned into it: its announcement says the model is "unofficially ML4, very officially: le Chonk."
Mistral Large 4 price and the launch discount
Right now Large 4 is only available through Mistral's paid developer service, Mistral Studio, where you pay for the amount of text it reads and writes. Prices are in US dollars, and Mistral has not announced any India-specific pricing.
- Standard price: $1.36 per million input tokens and $4.18 per million output tokens (a token is roughly three quarters of an English word)
- Launch offer: 50% off for the first two weeks, which brings it to $0.68 input and $2.09 output
- Previous model, Mistral Large 3: $0.50 input and $1.50 output, so the new one costs close to three times as much at full price
Artificial Analysis points out that in its tests, Large 4 costs more than four times as much per task as open models of similar ability from China, such as DeepSeek V4.1 Flash.
How to try Mistral Large 4 today
There isn't an easy free option yet. We couldn't find anything from Mistral or launch coverage saying Large 4 has arrived in Le Chat, Mistral's free ChatGPT-style app. These are your realistic options:
- Developer API (paid): sign up for Mistral Studio, add billing and pick the model called mistral-large-4. This is the only official way to use it today, and the 50% discount makes it cheapest in the first two weeks after launch
- Watch Le Chat: Le Chat is free to use and currently runs Mistral's older models. If Large 4 appears there, that will be the simplest way for non-developers to try it
- Download the weights (free, later): Mistral says the weights "drop end of this month". VentureBeat reports 27 October as the date, after about three weeks of testing with security experts and governments. VentureBeat also says the weights will use a custom Mistral licence, but Mistral hasn't published the terms yet
Mistral vs ChatGPT and Gemini: close, but not top of the table
Mistral's pitch is that Large 4 is the best open model from the US or Europe. It doesn't claim to beat the best closed models from OpenAI, Google or Anthropic overall.
Independent numbers back that up. Artificial Analysis gives Large 4 a score of 38 on its Intelligence Index, level with OpenAI's GPT-6 Luna and just behind DeepSeek V4.1 Flash on 39. That is a huge jump from Large 3, which scored 9 according to Simon Willison. Willison describes it as "maybe about 6 months behind the frontier".
Mistral's own tests show it winning in a few specific areas:
- Reading images closely: 42% on the Dense 200 visual test, just ahead of OpenAI's GPT-6-Astra on 41%
- Legal and finance tasks: Mistral says it beats GPT-6-Astra on a legal agent benchmark from Harvey and on Finance Agent v2
- Cybersecurity: 93% on Cybench, which Mistral says is the best score from any open model
- Coding: in human ratings by Surge AI, it came second out of five models, behind Anthropic's Claude Opus 5
Keep in mind that these are the company's own tests, and VentureBeat notes some competitor figures haven't been independently checked. Mistral didn't publish comparisons against Google's Gemini, so we can't fairly say how the two match up yet.
Why Europe cares so much about this model
Most of the strongest open models now come from Chinese labs such as DeepSeek, Qwen and Moonshot (which makes Kimi). CNBC frames Large 4 as Europe's attempt to close that gap. Mistral also says its European deployment runs separately under European law, which matters to banks, governments and other organisations that don't want sensitive data leaving the region.
Mistral is also leaning hard into security work. Its announcement argues that "defending software often starts with proving that a flaw is real", which is the kind of request closed models often refuse. Vice president Pierre Stock told TechCrunch that Mistral is working with governments and trusted partners to make sure the weights are used for defence and not for attacks.
What to watch next
The main date is the end of October, when the weights are due. VentureBeat reports 27 October. When they arrive, check the licence terms and see whether services you already use start offering Large 4 at lower prices than Mistral. Also watch for Large 4 showing up in Le Chat, which would make it the easiest way for everyday users to compare it with ChatGPT or Gemini side by side.
If you do try it, your prompts make a big difference to the answers you get. Our AI Prompt Playbook has prompts that work across ChatGPT, Gemini and Mistral.
Sources
- Introducing Mistral Large 4 (Mistral AI)
- Mistral Large 4 model card (Mistral Docs)
- Changelog (Mistral Docs)
- Mistral has released Mistral Large 4, making France home to the most intelligent model outside the US and China (Artificial Analysis)
- Mistral debuts Large 4 'Le Chonk' (VentureBeat)
- Mistral's new 1T model aims to leapfrog closed and open rivals (TechCrunch)
- Le Chonk (Simon Willison)
- Mistral unveils new AI model it says rivals best open systems from China (CNBC)
Found this useful? Send it to someone who should read it.


