Mistral Large 4: Open-Weight Cyber Pitch Faces Hard Questions

What Mistral launched

French AI lab Mistral has previewed Mistral Large 4, a very large model it has nicknamed "Le Chonk." The company is positioning it as an open-weight option for cybersecurity work that enterprises and governments can control themselves. 12 Mistral describes it as a natively multimodal model with roughly 1 trillion total parameters and 49 billion active parameters. 3 One analysis puts the figure at 1.05 trillion, which would make it the largest open-weight model from a European lab once the weights ship. 4

Those weights are not available yet. For now, Large 4 is offered only through a moderated API. Mistral says it will release the weights by the end of October, and several reports give October 27 as the date. 1345 Until then, Mistral says it is red-teaming the model with cybersecurity firms, vetted partners and state authorities. Those groups get access to a version with reduced moderation and expanded cyber capabilities. 1

The sovereignty argument

Mistral's case rests on control. The company argues that refusals built into a provider's model can block legitimate vulnerability research and incident response. It also argues that losing access to a capability in the middle of an incident is a security risk in its own right. 1 ZDNET describes the main selling point as security that stays under the user's control, unlike proprietary models whose access frontier labs can technically revoke. 2 Co-founder Guillaume Lample framed the model as a defensive tool. He said its cyber capabilities would help organizations fight threat actors who jailbreak closed models to carry out attacks. 2

On performance, Mistral says Large 4 is competitive with the strongest open models worldwide. It also claims the model clearly outperforms any open-weight model built in the US or Europe and is state-of-the-art among open models for cybersecurity, finance and law. 1 The company reports 82% on one Artificial Analysis Cyber Index test and 93% of the 40 Cybench security challenges solved. These are company-reported numbers. 3

Where the story gets complicated

Independent testing paints a less flattering picture. Artificial Analysis reportedly gives the model an Intelligence Index score of 38, ranking it somewhere between 25th and 64th among tracked models and behind GLM-5.3 and Kimi K3. Vals.ai places it 32nd of 44 overall. 5 The same analysis argues that the cybersecurity lead is narrower than Mistral's broader framing suggests. On reproduce-and-patch tasks, it says, Large 4's 82% partly reflects closed rivals refusing those prompts on policy grounds, not necessarily lacking the ability. 5

That point matters. If a benchmark advantage comes mainly from a willingness to answer, then the thing being sold is a moderation policy, not raw capability. Here that policy is "your policies, not ours."

The tension inside the pitch

The same feature Mistral markets as a strength is also the safety concern. RuntimeWire notes that letting customers run cyber work under their own policies also makes it easier to strip out safeguards once the weights are public. Independent testing and real deployments will show how that trade-off plays out. 3 Mistral's own rollout supports this reading. The company is staging the release through a moderated API while giving trusted parties a less restricted, more cyber-capable version. 13 That approach implies Mistral sees the unmoderated model as something that needs gatekeeping. Once the weights are downloadable, that gatekeeping ends, and the vetted-partner arrangement does not carry over to everyone else.

This is not necessarily a fatal contradiction. Defenders and attackers have always shared tools. But it does weaken the idea that open weights are simply the more secure choice. The argument that open weights put control in the user's hands applies equally to users with bad intentions.

Open, with caveats

The "open" label raises its own questions. Byteiota reports that every earlier Mistral model, including Mistral 7B, Mixtral 8x7B and Large 3, shipped under Apache 2.0. Large 4 is instead expected under a custom Mistral license whose terms have not been published. Such licenses can restrict commercial use, fine-tuning or third-party hosting. 4 Until those terms appear, claims about self-deployment and sovereignty remain unverified. 5

Hardware is another practical limit on who can actually self-host. Running Large 4 in FP8 reportedly takes about 1.26 TB of GPU memory, or eight H200 cards. That costs roughly $40,000 a month on AWS at continuous use. Even 4-bit quantization still requires around 630 GB of VRAM. 4 For most organizations, "you control it" will in practice mean renting expensive infrastructure or relying on a hosting partner.

The takeaway

Large 4 is a meaningful European entry, and its cybersecurity framing addresses a real frustration with closed models refusing legitimate security work. Still, the evidence so far points to a narrower story than the marketing suggests:

• Performance: general benchmarks are middling, and the cyber lead depends partly on a willingness to answer prompts rivals refuse.
• Safety: Mistral's own staged rollout implies it views the less moderated model as needing restriction, which open weights will make impossible.
• Openness: the license terms are still unknown.

The October 27 release, and the license that comes with it, will show whether the open-weight security pitch holds up.