Mistral gave its biggest AI model a nickname that sounds like a chunky cartoon cat. The numbers behind it are not cute at all.
One trillion parameters. Only 49 billion of them working at any moment. And model files you will be able to download on October 27.
The French company opened a public preview of Mistral Large 4 on October 6, 2026, and people have been arguing about it ever since. Some call it Europe’s big comeback. Others call it late and pricey. Both camps have a point, which is exactly why I wrote this post.
I went through the official announcement, the early press coverage and the independent scores. Then I sorted what is proven from what is only claimed. Short paragraphs, plain English, no hype.
By the end you will know what Le Chonk is, how it works, what it costs, where it shines, where it struggles, and what to check before the weights arrive in less than two weeks.
Table of Contents
- What Is Mistral Large 4, and Why Is It Called “Le Chonk”?
- One Trillion Parameters, Only 49 Billion Active: How Is That Possible?
- The October 27 Open Weights Date and the Three Week Test Window
- Trained From Scratch on About 4,000 Nvidia Chips in Europe
- The Benchmark Scores: Mistral’s Claims vs Independent Results
- The Cybersecurity Pitch and the Refusal Puzzle
- The Price: Cheap per Token, Not Always Cheap per Task
- Why You Cannot Run Le Chonk on a Laptop
- The License Question Nobody Has Answered Yet
- Weak Spots, Critics and a Builder’s Checklist Before October 27
1. What Is Mistral Large 4, and Why Is It Called “Le Chonk”?
Mistral Large 4, shortened to ML4, is the new flagship model from Mistral AI, the Paris based AI company. The official announcement describes a giant model built for coding, cybersecurity, finance, manufacturing and reading images.
It takes in text and images and answers in text. It does not draw pictures or generate audio, so do not expect it to replace an image generator.
It is also built for the world, not just for French speakers. The company says it was trained on more than 160 languages, and that list includes every official language of the European Union, according to Yahoo Tech’s launch report.
Now the name. “Le Chonk” started life as a joke. Before the model existed, a fat cat name floated around the AI community as a gag about Mistral’s missing trillion parameter model. The company later shipped a real trillion parameter flagship and kept the meme alive, as VentureBeat’s report on the release explains.
A serious company with a silly nickname is a smart move. People remember “Le Chonk” far faster than “ML4”.
There is a bigger story behind the joke, though. Mistral is pitching this model as the strongest open-weight model built outside China. That is a very specific claim, and it tells you who Mistral thinks it is competing against.
Related read: The US and China are not the only players in this race, and our breakdown of how Trump’s AI strategy frames the China race shows why a European open model is landing at such a loaded moment.
Quick summary of Thing 1:
- Official name: Mistral Large 4 (ML4)
- Nickname: Le Chonk
- Maker: Mistral AI, based in Paris
- Inputs: text and images
- Output: text only
- Public preview started: October 6, 2026
2. One Trillion Parameters, Only 49 Billion Active: How Is That Possible?
This is the part that confuses almost everyone, so let me make it simple.
Parameters are the numbers inside an AI model that store what it has learned. More parameters usually means more capacity to learn. One trillion is a huge number.
But Le Chonk is a “mixture of experts” model, often shortened to MoE. Think of a giant hospital with thousands of specialists. When you walk in, a front desk (the router) sends you to just a few of them, not to every doctor in the building.
Le Chonk works the same way. The network holds many small expert sections, and for each piece of text, the router wakes up only a few of them. DataCamp’s explainer on Le Chonk puts the numbers at roughly 1 trillion total and 49 billion active per token.
| Measure | Number | What it means in plain English |
|---|---|---|
| Total parameters | About 1 trillion | The full size of the model |
| Active parameters | 49 billion | What actually does the work for each token |
| Share that is active | Roughly 5 percent | About 1 in every 20 parameters |
| Input types | Text and images | It can read pictures, not make them |
| Output type | Text | Answers, code, summaries |
Why does this matter to you?
Speed and cost. The work the model does per word follows the 49 billion active parameters, so it can respond faster and cost less than a dense model of the same total size.
There is a catch, and we will get to it in Thing 8. The memory needed still follows the full trillion, because every expert has to be loaded and ready in case the router picks it.
Related read: Curious how the big closed models compare on coding, price and speed? See our Claude Opus 5.5 vs GPT-6 Sol vs Grok 4.7 comparison for the other side of the market.
3. The October 27 Open Weights Date and the Three Week Test Window
Here is the date that matters: October 27, 2026.
“Open weights” means Mistral will let anyone download the trained model files and run them on their own servers. You are not stuck calling Mistral’s API and paying per request. That is the whole point of the release.
Right now, only a preview exists. According to The Next Web’s coverage, developers can reach the model through Mistral’s API today, and the company plans to spend the next three weeks testing it in real world conditions before the weights go out.
Who gets early access? Developers, cybersecurity professionals and government authorities, through Mistral Studio, according to the Yahoo Tech report. Some of those groups will get a version with fewer restrictions and more cyber features than the public API.
Here is the timeline at a glance:
| Date | What happens |
|---|---|
| October 6, 2026 | Public preview opens on Mistral’s API |
| October 6 to 27 | Roughly three weeks of real world testing |
| October 27, 2026 | Open weights planned for public download |
| Same time, per Mistral | More architecture details, more benchmarks and the safety testing write up |
Notice what is missing before October 27: independent proof. Until the files are public, nobody outside Mistral can fully check the claims on their own hardware.
So treat every headline number you read this month as “Mistral says” until the weights land.
Quick summary of Thing 3:
- Preview is live now, weights are not
- Weights are due October 27
- Testing under real conditions is happening in the meantime
4. Trained From Scratch on About 4,000 Nvidia Chips in Europe
Let’s talk about how Le Chonk was made, because it is a big part of Mistral’s sales pitch.
Mistral says it trained the model from scratch, using Nvidia Grace Blackwell chips sitting inside its own European data centers. The training run took about two months.
The chip count depends on who you ask. Yahoo Tech’s report says 4,000, while Mistral’s own page, as explainx noted in its breakdown, says 3,800. Same ballpark, and a good reminder to use the official number when you quote it.
The same breakdown says Large 4 is the first milestone funded by Mistral’s 3 billion euro Series D round. That money explains the scale. A run like this is not a side project.
Why does “from scratch” matter? Because distillation, where a smaller model learns from the answers of a bigger rival, is a hot topic right now. Mistral stresses that it did not take that shortcut, and Techstrong.ai’s coverage points out that some Chinese developers have been accused of distillation.
That is Mistral’s framing, and it is a marketing point as much as a technical one. Still, it fits the bigger European message: built here, trained here, hosted here.
Related read: Thousands of Blackwell chips do not run on good wishes. Our report on how AI data centers are pushing up electricity bills explains the power demand behind runs like this one.
Quick summary of Thing 4:
- Trained from scratch, not distilled, says Mistral
- About 3,800 to 4,000 Nvidia Grace Blackwell GPUs
- Roughly two months of training in European data centers
- Funded by the 3 billion euro Series D
5. The Benchmark Scores: Mistral’s Claims vs Independent Results
Now the part everyone argues about. Are the scores good?
The honest answer is “good, but not the headline you may have read.”
Start with what Mistral claims. It says Le Chonk is the strongest open-weight model built outside China. It also says the model does well at agentic coding, scoring about 62 percent on DeepSWE, according to Technology.org’s launch summary. The same piece notes that Mistral itself calls its figures preliminary and expects them to shift before October 27.
Now the independent scores, which matter more.
Artificial Analysis rates the preview at 38.4 on its Intelligence Index, tied with OpenAI’s GPT-6 Luna. OfficeChai’s report calls that the highest score for a Western open-weights model on the chart. That is a giant leap from Mistral Large 3, which scored just 9.
But there is a twist. Trending Topics notes that Le Chonk is only eighth among all open models, behind Xiaomi’s MiMo-V2.6-Pro at 46.3, Z.ai’s GLM-5.3 at about 45 and Moonshot’s Kimi K3 at about 44. Closed leaders such as Claude Opus 5.5, at 57.6, are still well ahead.
The same report adds a cost note. Running the index cost about $1.13 per task, compared with about $0.27 for DeepSeek V4.1 Flash. Le Chonk also used roughly 200 million output tokens across the index, against a median of 81 million for other models.
Here is everything side by side:
| Test | Result | Who reported it |
|---|---|---|
| DeepSWE (coding agents) | About 62 percent | Mistral (vendor claim) |
| Cybench (cyber challenges) | 93 percent | Mistral (vendor claim) |
| Artificial Analysis Intelligence Index | 38.4, tied with GPT-6 Luna | Independent |
| Rank among open models | Eighth | Independent |
| Vals Index | 48.05 percent, 32nd of 44 models, ninth among open-weight | Independent |
| Harvey legal agent benchmark | First among open-weight models | Independent |
So the fair summary is this. Le Chonk is the best Western open-weight model on that chart, but it is not the best open model overall, and it is not close to the top closed models.
Mistral never claimed otherwise. The trouble starts when social media posts shorten “best outside China” into “best open model”, which is a different claim.
Quick summary of Thing 5:
- Best Western open-weight model on Artificial Analysis
- Eighth among all open models
- Far behind the top closed models
- Cost per task is higher than several cheaper rivals
6. The Cybersecurity Pitch and the Refusal Puzzle
Of everything in the launch, the cybersecurity angle is the most unusual.
Mistral says Le Chonk solves 93 percent of the Cybench challenges. It also says the model scores 82 percent on a test from the Artificial Analysis Cyber Index, which asks a model to reproduce a real open-source vulnerability and then patch it. According to AI Stock Wire’s report on the launch, Mistral also claims that Claude and GPT-6 refuse these cyber tasks, so they score near zero.
Read that again, because it is a bold claim. Mistral is saying the strict safety rules of the biggest closed models make them less useful for defenders who need to test their own systems.
Those scores are Mistral’s own numbers. Nobody outside the company has confirmed them yet.
Here is the puzzle. Mistral also says its model refuses malicious cyber requests more often than any other open model. A model that does more cyber work than rivals and also refuses more bad requests is a tricky balance to claim. Both statements can be true, but they sit awkwardly together until independent testers get hands on the weights.
As covered in Thing 3, a version with fewer restrictions and more cyber features goes to vetted groups first. That is a sensible rollout.
The real test arrives on October 27. Once the weights are public, anyone can fine-tune a model, and fine-tuning can weaken safety behavior. How Le Chonk’s guardrails hold up after that will be one of the most watched questions in AI security.
Related read: If AI and security worry you, our story on an AI agent that escaped its sandbox shows why model safeguards get so much attention.
Quick summary of Thing 6:
- 93 percent on Cybench, per Mistral
- Mistral says rival closed models refuse these tasks
- Mistral also claims high refusal of malicious prompts
- Independent proof has to wait for October 27
7. The Price: Cheap per Token, Not Always Cheap per Task
Let’s talk money, because this is where many readers get surprised.
According to Mistral’s own model page, the list price is $1.36 per million input tokens, $0.14 per million cached input tokens and $4.18 per million output tokens. The page also shows a promotional price at half those numbers, and commenters say that discount runs for roughly two weeks.
| Token type | List price per million | Promo price shown |
|---|---|---|
| Input | $1.36 | $0.68 |
| Cached input | $0.14 | $0.07 |
| Output | $4.18 | $2.09 |
That looks cheap next to many closed frontier models. It is not the cheapest open option though, and a price per token only tells half the story.
Remember the numbers from Thing 5. If a model talks a lot to finish a job, you pay for every extra word. That is how a low price per token can still become a high price per task, at about $1.13 per task on that index.
My advice is simple. Do not compare price lists. Run 30 to 50 of your own real tasks, then compare the final cost per solved task, not the sticker price.
Prices may also change once the preview ends, so check the docs page again before you build anything.
Related read: Price wars like this are part of a much larger money story, which we cover in our look at the AI bubble numbers behind OpenAI, Anthropic and SpaceXAI.
Quick summary of Thing 7:
- List price: $1.36 input, $4.18 output per million tokens
- Promo half price may be short lived
- Cost per task matters more than cost per token
8. Why You Cannot Run Le Chonk on a Laptop?
“Open weights” sounds like you can grab the file and run it at home. For this model, that is not realistic.
Remember the memory catch from Thing 2. Speed follows the 49 billion active parameters, but memory follows the full trillion. Every expert must sit in GPU memory, because the router might call on any of them.
Simple math shows the problem. One trillion parameters at 8-bit precision is roughly 1 terabyte of weights. At 4-bit precision it is about half that. A gaming PC has a tiny fraction of that space, and even a good workstation falls far short. This technical deep dive on running Le Chonk locally walks through what the hardware would look like.
So who is the open release really for?
- Banks, ministries and manufacturers that must keep data inside their own walls or inside Europe.
- Hosting companies that can offer the model at competitive prices once the files are public.
- Researchers and fine-tuners who want to inspect it, shrink it, or adapt it.
If you are a solo creator or small developer, the benefit reaches you indirectly. More companies hosting the same model usually means better prices and more choice.
Quick summary of Thing 8:
- Compute follows 49 billion, memory follows 1 trillion
- You need a GPU cluster, not a laptop
- The big winners are regulated and sovereign buyers
9. The License Question Nobody Has Answered Yet
This might be the most important detail on the whole list, and it is still a blank.
Open weights does not automatically mean “do anything you like.” VentureBeat’s report mentioned earlier says the model will use a custom Mistral license, and the full terms arrive together with the weights on October 27.
A custom license can hide surprises. Before you build a product on Le Chonk, check these points on day one:
- Can you use it commercially, and are there company size or revenue limits?
- Can you redistribute the weights?
- What happens to models you fine-tune from it?
- Are there banned uses?
Until those answers are public, “open” is a promise, not a guarantee. Mistral also said it would publish its architecture details, extra benchmarks and the safety testing on the same day, which gives the community plenty to verify.
That kind of outside checking is exactly why people want models tested by rivals and neutral parties, an idea we explored when Elon Musk proposed that AI labs test each other’s models.
Quick summary of Thing 9:
- License is reported as custom, not a standard open license
- Full terms arrive with the weights
- Read it before you build anything commercial
10. Weak Spots, Critics and a Builder’s Checklist Before October 27
No model is perfect, and Mistral is honest about at least one weakness. It admits Le Chonk still trails other frontier models in coding, according to the Yahoo Tech report.
Critics have said more. Some developers on Hacker News, as summarized in the explainx breakdown, called the model late and only average, and complained about the cost per task. Others noticed that some launch charts were arranged in ways that made Mistral look better. That is common in AI launches, and it is why independent testing matters.
Fans have their own list of reasons to like it. They love the open weights, the European hosting and the focus on security work. And the jump from Mistral Large 3 is real.
So what should you actually do in the days before October 27? Here is a simple plan.
- Test with your own tasks. Take 30 to 50 real jobs from your work and run them through the preview API. Track quality, speed and cost.
- Try your images. If you handle PDFs, charts or photos, test the vision features yourself, since Mistral highlights image understanding.
- Test long inputs. The docs list a large context window, so check how it behaves on your longest files.
- Decide where you would host it. Mistral’s API, a third party host or your own cluster all have different data and cost trade-offs.
- Read the license the day it appears. Then decide.
Planning to use it for agents that act on your behalf? Think carefully about permissions first, and read our guide on what you should never hand over to an AI agent.
Quick summary of Thing 10:
- Mistral admits it trails in coding
- Critics call it late and costly per task
- Run your own tests instead of trusting charts
Mistral Large 4 FAQ
What is Mistral Large 4, or Le Chonk?
It is Mistral AI’s flagship model, a mixture of experts system with about 1 trillion parameters and 49 billion active per token. It reads text and images and answers in text.
When can I download the Le Chonk weights?
Mistral plans to release the weights on October 27, 2026, after a roughly three week test period.
Is Mistral Large 4 really open source?
It is described as open weights, and the license is reported to be custom, not a standard open license. Check the full terms when they are published.
Is Le Chonk better than Claude or GPT?
Not overall. Independent scores place it as the top Western open-weight model on one chart, but closed leaders such as Claude Opus 5.5 score higher.
Can I use it today?
Yes, through Mistral’s API and Mistral Studio during the preview, though some versions are reserved for vetted groups.
The Bottom Line
Le Chonk is a real release with real numbers, wrapped in a funny nickname. It is strong, but it is not the best model in the world, and Mistral does not claim it is.
The honest story is about choice. Europe now has a trillion parameter open-weight model, trained on its own soil, aimed at companies that want control over their data.
Whether it earns a spot in your toolkit depends on three things: the independent tests, the license, and your own results. All three become much clearer on October 27.
Until then, test it yourself and trust your own numbers over any launch chart.