Mistral AI published the open weights of Mistral Large 4 on Thursday, October 29, putting a 1-trillion-parameter model on Hugging Face where anyone can download it. The company aimed the release at a narrow audience: cybersecurity teams that want to run vulnerability work on their own hardware. It landed one day after Google kept its newest cyber-capable model, Gemini 4 Carbon, limited to vetted defenders, which makes the open-versus-closed debate over security AI concrete this week.
This is not a consumer product, and Mistral is not asking the public to chat with it. The more pressing question is who controls a model like this once it exists, and the answer is less comfortable than the launch materials suggest.
What Mistral released, and when
Mistral published a preview of Mistral Large 4 on October 6, then released the full weights about three weeks later. The company calls the model "Le Chonk" and describes it as a multimodal system with 1 trillion parameters, according to its announcement and technical introduction. A third-party summary from Data Studios describes a mixture-of-experts design with about 49 billion active parameters and a 1-million-token context window.
The weights sit on Hugging Face under a modified MIT-style license. Mistral calls it the Mistral Ecosystem Open-Weight License, and it adds limits on offering the model as a hosted service. According to Mistral's announcement, as reported by The420, the gains after the preview came from additional training and testing with government agencies and corporate users.
Mistral says the model is built to help organizations find and fix software vulnerabilities, and it is pitched at teams that want to run that work on their own infrastructure. The company also says it is the strongest open-weight model it has built and that it significantly outperforms any open-weight model developed in the US or Europe. Hugging Face's writeup ranks the model fifth among open-weight models on the Artificial Analysis Intelligence Index.
Why Mistral says open weights matter for defenders
Mistral's argument is operational rather than academic. Provider-level refusals can block legitimate vulnerability research and incident response, the company says, and losing access to a capability in the middle of an incident can itself be a security risk. For a team working a live intrusion, a refused request is an operational problem, and a self-hosted model takes the provider out of that loop.
The release also carries a sovereignty pitch. Mistral is a French company, and it positions the model as a self-hosted alternative to closed US systems from OpenAI, Google and Anthropic. Orange Cyberdefense, the cybersecurity arm of France's Orange, has partnered with Mistral to put a cyber-focused version of the model into selected services, according to Reuters as reported by Tech Startups.
How it compares with Google's restricted Gemini 4 Carbon
The timing makes the comparison unavoidable. Google released Gemini 4 Carbon first to a subset of vetted cyber defenders, with broader access planned gradually through its Fairwind Program. WinBuzzer reports that Gemini 4 Carbon scores 59 on the Artificial Analysis index, one point above Claude Opus 5.5. Google's introductory API price is $2 per million input tokens and $10 per million output tokens, rising to $4 and $20, according to Anadolu Agency. USA 24 has covered those access rules separately in its Gemini 4 Carbon report.
The structural difference is simple. Google decides who can call its model and can pull that access back. Once Mistral's weights are on someone else's disk, Mistral has no such lever. Pierre Stock, identified in coverage as a Mistral executive, reportedly acknowledged in reporting summarized by Shattered that once the weights are public, access cannot be contained and fine-tuning can change behavior.
Mistral's comparison set is narrow, too. Its "best open-weight" framing measures the model only against other open models, not closed frontier systems. That is a defensible claim on its own terms, but it is different from leading all of cyber AI, and readers should weigh the two separately.
Which benchmark numbers are verified, and which are not
Mistral and secondary coverage cite several figures. Nearly all of them are vendor-reported, and the table below separates them from third-party rankings.
| Claim | Figure | Source | Status |
|---|---|---|---|
| Artificial Analysis Cyber Index | 51% | Shattered | Vendor-reported |
| CyberGym-E2E | 79–82% | Crypto Briefing | Vendor-reported |
| Reproducing and patching a real open-source vulnerability | 82% | Shattered, citing Mistral | Vendor-reported, not independently audited |
| Terminal-Bench 4.0 | 41%, up 13 points from preview | Mistral | Vendor-reported |
| Open-weight ranking, Artificial Analysis Intelligence Index | Fifth | Hugging Face | Third-party ranking |
Two details call for caution. The 82% patching figure comes from a Mistral announcement as reported by a third-party site, and it has not been independently audited. The Terminal-Bench 4.0 score of 41% is up 13 points from the preview, according to Mistral. As of this writing, independent verification of the cyber benchmarks is still missing.
Hugging Face says the model's refusal rate is higher than that of any other open-weight model. The exact figure has not been confirmed, and a specific refusal percentage circulating on lower-quality sites has not been verified.
License terms, API status and the price question
The license is the first document a company should read. The Mistral Ecosystem Open-Weight License is a modified MIT license that limits model-as-a-service offerings. A business that wants to resell access to the model as a hosted service needs to check those limits before building anything on top of it.
The hosted API is less settled. Mistral's API version still carries a "public preview" label on its model card, so general-availability and deprecation terms are not yet set. A third-party site cites $0.70 per million input tokens and $2.10 per million output tokens, but Shattered reports that Mistral's release materials do not confirm a price. Treat the price as unverified until Mistral's pricing page is checked.
So is it free? The weights can be downloaded, and the license governs what you can do with them. Running a model of this size on your own hardware is a separate cost, and none of the coverage quantifies it.
Who gains and who loses from open weights
Security teams and enterprises that want to run cyber tools without sending sensitive data to a cloud vendor now have a near-frontier option they can host themselves, according to coverage from Tech Startups and Data Studios. For organizations with strict data rules, that changes the procurement conversation.
Misuse controls move, too. Guardrails set on Mistral's API do not reach copies that others download and run, so responsibility for misuse shifts toward whoever deploys the model. The concern is not theoretical for the industry. Anthropic's IPO filing, covered by USA 24, discloses models that reached real systems, which is a reminder that deployment controls matter as much as model design.
Policymakers face the harder problem. Once downloaded, the model cannot be recalled or fully contained. A letter from 23 state attorneys general, plus DC and American Samoa, asked Congress for federal AI regulation and safety protocols, warning of risks to national security and infrastructure, as reported by ABC News. That letter is about five weeks old and is background only. No official government response to this release has been found, and whether US agencies will adopt or restrict the model is unknown.
What security teams should do before testing it
Non-technical readers have no direct action to take, because the model is not a consumer product. Security teams evaluating it should start with these steps.
- Read the Mistral Ecosystem Open-Weight License in full, including its limits on hosted services.
- Download the weights from the official Hugging Face repository, and verify file integrity before loading them into your environment.
- Test refusal behavior on your own prompts instead of relying on vendor claims.
- Re-run the tasks your team actually needs. The 82% patching figure has not been independently audited.
- Check Mistral's pricing page before budgeting for the API, since the $0.70 and $2.10 figures are unconfirmed.
- Restrict who can fine-tune the model and log changes, because fine-tuning can change behavior.
What to watch next
Several open questions will decide how much this release matters. Independent evaluators need to test the cyber numbers. Mistral needs to clarify API pricing and move the model out of public preview. US agencies need to say whether they will use or restrict it.
Google's Fairwind rollout will show whether restricted access can hold when a model with strong cyber claims is one download away. The state attorneys general asked Congress to act, and a model anyone can download raises the stakes of that request. The next test is whether defenders who adopt Mistral's model see the results Mistral describes.
