Here’s big news from France. Mistral AI has just released the full open weights of Mistral Large 4, its 1-trillion-parameter model. It happened on October 29, 2026. Now anyone can download it, study it, and run it on their own systems. And the new version codes better than the preview did.
What happened
So, what actually changed on October 29? French AI company Mistral AI released the weights of Mistral Large 4. Fans call the model Le Chonk.
This was the step a lot of people had been waiting for. Back on October 6, Mistral only opened a preview. You could try the model through its API, sure, but you couldn’t download it. That has changed now. The weights are live on Hugging Face, and developers can grab the model files and run them on their own servers.
Mistral says it added three more weeks of extra training since the preview. It took in feedback from early users, too. The result, it says, is a much stronger model for coding and security work. And the gains, Mistral claims, are still growing.
If you missed our first story, here’s the link to our earlier post on the Mistral Large 4 preview launch.
What is new since the preview
Let’s start with coding, because that’s where the jump is clearest. Mistral says Large 4 now leads open models on several coding tests.
First, FrontierCode 2.0. This test checks full coding jobs, start to finish. The model scored 19.7%. Mistral says that’s the top score among open models.
Next, the Artificial Analysis Coding Agent Index. Here it scored 48%. That ties the leading open model, GLM-5.3. But here’s the catch: Mistral says Large 4 uses 47% fewer tokens per task on average. Fewer tokens mean a lower bill for the same job. That adds up fast if you run a lot of tasks.
Then there’s Terminal-Bench 4.0, for coding in a terminal. The score went from 28% in the preview to 41% now. That’s a jump of 13 points. Output tokens dropped by 30% at the same time. In plain words, it does more coding work and spends less doing it.
Mistral also added three reasoning modes. Low fits simple tasks. Medium fits everyday tasks. High is the default, and it fits hard tasks that need deep care. You pick a mode based on how hard your task is, and how much you want to spend.
Security work also improved
Security work got better, too. On the Artificial Analysis Cyber Index, the model scored 51%. What does that test check? Three skills. Can a model find a weak point in code? Can it copy the flaw in a safe test? And can it fix it without breaking the rest? Mistral also says the model refuses 96.6% of harmful cyber requests on average. That figure comes from tests with JailbreakBench, StrongREJECT, and AgentHarm.
Now, one honest warning, straight from Mistral. The company watched live use after the preview. It says the model refused direct attacks on live targets. But it still sees a risk in some cases. The model could help a real attack. Mistral says it is changing its safety rules to deal with this, and more updates will come in the next few months.
One more thing to know: a special version for deep security work is NOT open to all. Mistral gives it only to vetted business partners, such as Orange Cyberdefense. Their testing must stay on systems they own, or ones they have clear permission to test.
The size and the specs, in simple words
Large 4 has 1 trillion total parameters. Only about 52 billion are active for each task. Think of parameters as small settings the model learned during training. The full set is huge, but only a small part works on each job. That keeps the cost down.
The model reads text and images, and it writes text and code. It can take in up to 1 million tokens of context at once. That’s a very long input, close to 750,000 English words. It also works in more than 160 languages, including every official language of the European Union.
Mistral trained the model from scratch in Europe. It used 3,800 NVIDIA Grace Blackwell GPUs in its own data centers. The company says this gives users a clear view of where the model comes from, and how it was built.
The price
Through the Mistral API, Large 4 now costs $0.70 per million input tokens. Cached input costs $0.07 per million tokens. Output costs $2.10 per million tokens.
For most students and freelancers, the API is the easy route. You pay only for what you use. You don’t need to buy costly hardware.
You can also try the model in Mistral Studio, through OpenRouter, or in Mistral’s Vibe Code coding tool. NVIDIA offers hosted endpoints too. And Microsoft is the first cloud provider to offer the model through Microsoft Foundry.
The license rule you must know
The weights are free to use, change, fine-tune, and sell as part of your own work. The license is called the Mistral Ecosystem Open-Weight license. It’s a changed form of the MIT license.
There’s one key condition, though. If your business sells access to the model itself as a service, you need one more thing first. You need a separate commercial license from Mistral. This rule does not stop a freelancer from building an app on top of the model. It stops a company from reselling the model as its own paid service without a deal.
A smaller quantized version is also out for Blackwell GPUs. It needs less memory than the full model. Even so, the full model still needs serious hardware. Don’t expect to run it on a normal laptop.
Why this matters for freelancers
Freelancers who build AI tools now have a strong open choice. The API price is low. You can build a coding helper, a document reader, or a chatbot for a client. You do not need to pay top closed-model prices.
The open weights also help with privacy. Say a client wants their data to stay on their own servers. You can run the model for them there. Their data never has to go to a third-party API. That can win you work with careful clients, such as law firms and finance teams.
The three reasoning modes help you control cost, too. Use Low for simple jobs. Use High only for hard jobs. That way, you don’t pay top price for easy work.
Why this matters for students
Students get a free way to study a top-level model up close. You can read the model card. You can test the API at a low cost. And you can build class projects with a real frontier model.
If you want to learn AI coding, this release is a good chance to practise. Build a small coding agent. Test it on a real repo. Compare Low, Medium, and High modes, and see how the answers and costs change. That kind of hands-on work looks strong in a portfolio.
For more free tools to pair with this news, see our list of free AI tools for students.
What to watch next
First, watch independent tests. Most scores in this article come from Mistral’s own tests. Independent sites will now test the open weights themselves. Their results will show if the claims hold up in the real world.
Second, watch the price. The current API price is low, but prices can change. Check the live price in Mistral Studio before you start a big paid project.
Third, watch the safety updates. Mistral says it will keep updating its guardrails in the next few months. The rules for security work may change as those updates land.
The bottom line
The preview was a promise. The open weights release is the real thing. Mistral Large 4 is now free to download. It is stronger at coding than it was three weeks ago. And its API price is low.
If you are a freelancer or student, this release is for you. You get strong AI without big-company prices. It is one of the most useful releases of October 2026. Try it on a small project first, watch the costs, and build from there.
Sources
- Mistral AI’s official announcement (October 29, 2026)
- Reuters: Mistral AI publicly releases Large 4 (October 29, 2026)
- Unite.AI: Mistral’s Le Chonk checkpoint lands on Hugging Face (October 29, 2026)