On October 29, 2026, the French company Mistral AI publicly released Mistral Large 4, its largest model, with its full files free to download. The short answer for a Saudi business owner: you can run it on your own servers, keep customer data inside your company, and use it commercially without paying a license fee, as long as you do not resell the model itself as an AI service. The catch is hardware: the official setup uses 8 data-center GPUs. Most companies should start with the API at $0.70 per million input tokens (about 2.63 riyals) and $2.10 per million output tokens (about 7.88 riyals), then move in-house only when the volume or the sensitivity of the data justifies it.
What Mistral released, in plain terms
Mistral Large 4 has about 1 trillion parameters, but it uses only around 52 billion of them for each piece of text it produces, which keeps it faster and cheaper to run than its size suggests. It reads text and images, can hold up to 1 million tokens in one request (a long contract with all its annexes fits comfortably), and Mistral says it is fluent in more than 160 languages. It entered a limited preview with selected partners on October 6, 2026, and the full public release followed on October 29.
The phrase «open weights» means you receive the model itself, not just access to it. You can install it on servers you control, inside the Kingdom if you choose, and no request ever has to leave your network. If you read our coverage of Inkling from Thinking Machines, the idea is the same. What is new here is the license, the backing of an established European vendor, and the range of places you can run it.
The one condition in the license
The model ships under the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT license. You may use it, modify it, fine-tune it on your own data and use it commercially. The one condition: a business that runs a Model-as-a-Service offering, meaning it sells access to the model itself to other companies, must first obtain a separate commercial license from Mistral.
For most Saudi businesses this condition does not apply. A clinic that uses the model to summarize patient files, a distributor that uses it to answer customer messages, or a law firm that uses it to review contracts is using the model inside its own work, not selling it. If you are a software company planning to offer an AI service built on top of it to your own clients, read the license with your lawyer before you launch.
Three ways to use it, and what each one costs
- Mistral's API: $0.70 per million input tokens, $0.07 for cached input and $2.10 per million output tokens. You pay per use and manage no servers, but your data travels to Mistral.
- A cloud platform: Mistral says Microsoft Foundry is the first cloud provider to offer the model, and it is also available through OpenRouter and NVIDIA's hosted endpoints. This suits a company that already has an account and a contract with one of these providers.
- Your own servers: The weights are on Hugging Face. The official deployment examples run on 8 NVIDIA GPUs working together, and a lighter compressed version targets NVIDIA's newer Blackwell chips. You pay for the hardware or the rented servers, not per request.
An honest note on cost: when Artificial Analysis tested the preview, Fortune reports, it rated the model roughly on par with OpenAI's GPT-6 Luna but found it cost more than four times as much per task as open models of similar intelligence. Price alone is therefore not the reason to choose it.
Why owning the model matters: two real risks
Data leaving the Kingdom. When you send customers' personal data to an AI provider abroad, that transfer falls under the Saudi Data and AI Authority's Regulation on Personal Data Transfer Outside the Kingdom, which relies on measures such as appropriate safeguards and standard contractual clauses. Running the model on servers inside the Kingdom takes that transfer out of the picture for this use, although every other obligation of the Personal Data Protection Law still applies. Our article on cloud regions inside the Kingdom explains what hosting location does and does not solve.
Access being cut off. In June 2026, a US government directive forced Anthropic to suspend access to its Fable 5 and Mythos 5 models for nearly three weeks, as Fortune recalls. We wrote at the time about what that meant for businesses built on a single provider. A copy of a model running on your own servers cannot be switched off from outside. Mistral's vice president of science, Pierre Stock, made exactly this argument at launch.
How strong is it, really?
The figures come from Mistral and should be read as such. The company says the model narrowly leads open models on the Artificial Analysis Coding Agent Index at 48%, scores 41% on Terminal-Bench 4.0 (up from 28% in the preview, a tie with China's GLM-5.3), and resists 97.1% of attacks on the B3 agent security benchmark. Fortune adds the context: closed models from OpenAI and Anthropic remain ahead overall; Anthropic's Mythos 5.1, for example, scores 60.9% on Terminal-Bench 4.0. In short, this is a serious open option, not the strongest model on the market.
When it makes sense for your business
- You handle sensitive data (health, finance, legal) and want it to stay on your own infrastructure.
- You have high, steady volume, where a fixed server cost beats a per-request bill.
- A core workflow cannot afford to stop because a provider changes its terms or loses access.
- You want to fine-tune a model on your own documents and own the result.
If none of these apply, a closed model through an API is usually simpler and stronger. Arabic quality is also unconfirmed: the official pages mention more than 160 languages without naming Arabic, so test the model on your real Arabic documents before you commit.
The Origami view
We do not advise buying 8 GPUs because a model became downloadable. We advise building your AI layer so the model can be swapped: start with an API, measure accuracy on your own Arabic and English data, keep the prompts and the data pipeline in your hands, and move sensitive workflows to a self-hosted model such as Mistral Large 4 when the numbers and the risk justify it. If you want help testing it on your data or designing that layer, talk to us.
Sources
Frequently asked questions
Is Mistral Large 4 free for my company to use?
The model files are free to download, and the license allows commercial use, modification and fine-tuning on your data. The only exception is selling access to the model itself as a service to other companies, which needs a separate license from Mistral. You still pay for servers if you host it, or pay per use through the API: $0.70 per million input tokens and $2.10 per million output tokens.
What hardware do I need to run Mistral Large 4 on my own servers?
The official deployment examples on Hugging Face use 8 NVIDIA GPUs working together, with a lighter compressed version for the newer Blackwell chips. It does not run on a desktop or an ordinary server, which is why most companies start with the API or a cloud platform.
Does running it inside the Kingdom cover Personal Data Protection Law compliance?
It removes the cross-border transfer for that use, which is the transfer governed by SDAIA's Regulation on Personal Data Transfer Outside the Kingdom. The law's other obligations, such as a lawful basis for processing, data subject rights and access controls, still apply.
Does Mistral Large 4 support Arabic?
Mistral says it is fluent in more than 160 languages, but its official pages do not name Arabic explicitly. Test it on real Arabic documents and messages from your business before relying on it.
