Mistral lets firms download Large 4 and run it themselves
The model reads text and images, works in more than 160 languages and is free to use commercially, except for businesses that offer the model to others as a service. API prices start at about 24 baht per million tokens.
Jump to a section
When a bank or a hospital wants AI to hunt for holes in its systems, the hard part is not only how good the AI is. It is that the code and customer data have to be sent to another company's servers to be read, and many organisations cannot do that.
That is the problem Mistral, the French AI company, is trying to answer. On 29 October 2026 it released the weights of Mistral Large 4 for anyone to download on Hugging Face, TechCrunch reported. The model has about 1 trillion parameters, and Mistral is pitching it as a cyber defence tool for companies and governments.
Files like these are called open weights: the full set of numbers the model learned. Once downloaded, an organisation can run the model on its own servers, and the data never has to leave the building. Pierre Stock, Mistral's vice president of science, told TechCrunch that because the model is open, nobody can cut defenders off from it.
1 trillion parameters, but 49 billion at work at a time
Parameters are the numbers a model adjusts as it learns. More of them usually means more detail, but also more computing power. So Mistral Large 4 uses a design called Mixture-of-Experts.
Picture a hospital with 128 specialists. Each patient does not see every doctor, only the 4 who fit the symptoms, plus 1 house doctor. Mistral's model card says the model picks 4 of 128 experts for each token, plus 1 shared expert, so about 49 billion parameters are at work, or 52 billion counting the input and output layers.
The model reads text and images, answers in text, handles a context of 1 million tokens and works in more than 160 languages. It is not a laptop model: the examples in the model card run on 8 GPUs, and a compressed NVFP4 version is built for Nvidia's Blackwell GPUs.
Strong at cyber defence, and Mistral admits the risk
The numbers in this section are Mistral's own, from the company's announcement. It said the model refuses malicious cyber requests 96.6% of the time on average, the highest among the open models it measures, and resists 97.1% of attacks on the B3 Agent Security benchmark, up from 93.3% for the preview.
On coding, the model scored 48% on Artificial Analysis' Coding Agent Index, close to GLM-5.3, a Chinese open model, while using 47% fewer tokens per task. TechCrunch reported that figure as Mistral's claim.
The notable part is that Mistral said plainly that after the preview it saw potential for the model to help people carry out real attacks, even though in testing the model refused black-box attacks on live targets. It is tightening its guardrails as a result. A specialised cyber version is open only to partners such as Orange Cyberdefense, after vetting and contracts, and may be tested only on their own systems or systems they are authorised to test.
Free to use for profit, but a hosted service needs a licence
Mistral uses a licence called the Mistral Ecosystem Open-Weight licence, a modified MIT licence. Organisations can use the model, change it, fine-tune it and make money from it. There is one condition: a business that offers the model for others to call as a Model-as-a-Service needs a separate commercial licence from Mistral.
Teams without the hardware can use the API. Mistral's documentation lists $0.70 per million input tokens, or about 24 baht, and $2.10 per million output tokens, or about 71 baht, at 33.69 baht to the dollar on 2 November 2026, according to Yahoo Finance. Cached input costs $0.07 per million tokens. Besides Mistral's own API, the model is available through OpenRouter, Nvidia and Microsoft Foundry.
How Thai organisations can try it
- Start with the API and a real task from your team, such as asking it to read part of your code and flag risky spots, because every score so far is Mistral's own test
- To run it yourself, plan for servers with data centre GPUs, as the model card's examples use 8
- Read the licence first if your plan is to offer the model to customers as a service
- Test security only on your own systems, or systems whose owner has given written permission
The weights have been on Hugging Face since 29 October 2026, and teams that want to try the API can start today.
FAQ
How are open weights different from a model you use through an app?
A model used through an app or API runs on the maker's servers. Open weights are files you download and run on your own machines, so your data does not have to be sent outside.
Can I run Mistral Large 4 on a personal computer?
The model card's examples run on 8 GPUs, so it suits servers rather than personal machines. Most people will find the API easier to try.
How far can the security scores be trusted?
They are Mistral's own test results, and the company is still adjusting its guardrails. Organisations should try it on their own work before relying on it.
How much does the API cost?
$0.70, or about 24 baht, per million input tokens and $2.10, or about 71 baht, per million output tokens, according to Mistral's documentation.
