Mistral Large 4 goes open-weight: what SMBs should check

1. What was released

In its official post, Mistral says it has released the weights of an improved Mistral Large 4 (nicknamed Le Chonk), the flagship model it first put out as a public preview earlier this month. It is a one-trillion-parameter multimodal model, and Mistral says three more weeks of reinforcement learning since the preview raised its coding and defensive-security performance (Mistral announcement). According to TechCrunch, the weights are available on Hugging Face and through Mistral's API, so companies and government agencies can download and run the model themselves (TechCrunch).

According to GIGAZINE, the model is a mixture-of-experts design with one trillion total parameters, 52 billion of them active at a time, and a context window of up to one million tokens. It is released under the Mistral Ecosystem Open-Weight License, which allows further training and commercial use, but companies that offer the model to third parties as an API service need a commercial licence (GIGAZINE). We think this condition will rarely get in the way of using the model for a company's own work, but check the licence text itself before committing.

2. How to read the scores

Mistral says Mistral Large 4 scores 48% on Artificial Analysis's Coding Agent Index, effectively tied with the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task on average. In security, it cites 51% on Artificial Analysis's Cyber Index, which measures finding, reproducing and patching vulnerabilities, and an average 96.6% refusal rate on malicious cyber requests, the highest it has measured among open models (all figures are Mistral's own account in its announcement).

GIGAZINE reports that on the same Coding Agent Index, Mistral Large 4 with Mistral Vibe scored 48, level with GPT-6.1 Terra with Codex, but that on the combined Intelligence Index v5.0 it scored 26, below DeepSeek V4.1 Flash 1017 and GLM-5.3 (GIGAZINE). In other words, it is strong at coding and security but is not the best at everything.

Our view For a small business, we think the point is less the ranking than getting an AI that cannot be switched off. Mistral's Pierre Stock says that with open weights, access cannot be cut off by the provider (TechCrunch). In industries where clients object to customer lists or quotation data being sent to an overseas API, a company with a choice and one limited to cloud AI will end up qualifying for different contracts. Running a one-trillion-parameter model in-house needs heavy equipment, so we expect the realistic route for small firms to be a provider that hosts this model in a Japanese or EU cloud. Listing your own data categories, what may leave the company and what may not, this month will speed up that decision.

3. What is moving in security

Mistral's announcement carries comments from several security companies. The EU agency ENISA says it is testing Mistral Large 4 to identify vulnerabilities, audit code and analyse defensive logs, and Elastic says Elastic Security already supports Mistral's open models. Zscaler says its Security for AI offerings now integrate with Mistral (all as quoted in Mistral's announcement; we have not checked each company's own pages).

Stock said that because AI is already helping find bugs much faster, the same capability needs to reach defenders quickly, and he acknowledged that attackers may use their own technology to attack (TechCrunch). We covered the case of AI's role in the critical PHP flaws in our October 25 brief.

4. What small businesses should check now

List the generative AI tools in use that receive customer details or quotation data, and write down what data may be entered
Check your contracts and client supplier standards for clauses that restrict where data is stored or whether it may be sent overseas
Ask vendors of AI-based business tools which model they use and where to find out which country the data is processed in
If you ask a Japanese cloud provider or developer to quote a private deployment of an open-weight model, confirm the licence terms (commercial use, offering to third parties) at the same time
If you use AI for security checks, decide first who will judge the results and act on them

Our outlook

Over the next six months to two years, we expect weights for near-frontier models to be published regularly, and whether a company rents AI as an outside service or runs it under its own control to come up more often as a condition of doing business. In the same period, AI that searches for and compares suppliers on behalf of consumers and buyers will spread; if it also checks data handling against the buyer's standards, companies that cannot explain how they handle data will be dropped at the comparison stage. For a small company, what pays off is not owning a giant model but classifying its data, putting its AI policy in writing, and building workflows in which the model can be swapped as circumstances change. Now is the time to lay that groundwork.

Questions this article answers

What does the release of Mistral Large 4's weights change for small businesses?

It widens the option of running AI in an environment you control rather than sending data to an outside API. A one-trillion-parameter model needs heavy equipment, though, so for most small businesses the realistic route is a provider that hosts it in a Japanese or EU cloud. Start by deciding what data may be entered.

Can open-weight models be used commercially without restriction?

Licences differ by model. According to GIGAZINE, Mistral Large 4's licence allows further training and commercial use, but companies that offer the model to third parties as an API service need a commercial licence. Always read the licence text before adopting a model.

Sources

Related Q&A

Related services

If you want to try this in your own business

To see how the topics in this article would apply to your own operations, and what it would cost, send us a message through the contact form.