ZDNET’s key takeaways
- Mistral’s open-weight ML4 model Le Chonk is in preview.
- ML4 was trained with fewer GPUs than OpenAI’s Astra, but competes.
- Open models are positioned as democratic defenders from AI attacks.
French AI lab Mistral has shipped its latest model, Mistral Large 4 (ML4) — and it’s positioned as the open solution for all your defense requirements.
Also: Open weights vs. closed: An AI civil war’s afoot, and the stakes are existential
One outcome of recent AI security incidents is that open and proprietary models are being pitted against each other. Initially framed as less safe because of their malleability, open models are viewed as a new option for cyber defense after proprietary models from Anthropic and OpenAI proved just as risky.
Mistral said ML4, which the company has nicknamed “Le Chonk” for its trillion-parameter size, is built for security that stays under user control — unlike proprietary models, which frontier labs can technically rescind access to at any time.
“The cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyberattacks,” Mistral co-founder Guillaume Lample said.
Le Chonk and security
After a hack-filled summer that put AI model security under the spotlight, everyone is looking for a reliable AI security solution that suits their needs. ML4 prioritizes cyber defense capabilities, advertising customizable control and data sovereignty.
Also: This new ChatGPT scam tricks you into installing malware – how to spot the trap
In a briefing, Lample and Mistral’s VP of Science Pierre Stock emphasized that security is a key requirement for the company’s enterprise clients, echoing an ongoing industry trend. Following the Hugging Face breach, Mistral was one of many companies that signed Nvidia’s Open Secure AI Alliance, a cross-industry partnership that argued open models are crucial to democratizing defenses against increasingly common AI security incidents.
“ML4 is the beginning of a leading generation of open-weight, customizable, cybersecurity models that enterprises can fully own and control, without vendor lock-in,” Mistral wrote. “Enterprises and states should not have to rely on a closed model vendor that could arbitrarily turn off their cyber defense capabilities.”
“In absolute terms on cyber capabilities, ML4 outperforms the best models from Kimi, Deepseek and Meta,” a Mistral spokesperson told ZDNET
Also: Who owns AI risk at work? Business and tech leaders can’t agree, PwC survey finds
Le Chonk is available now in public preview. Mistral said it will release the model weights on Oct. 27. That time gap gives the lab a month “to work with developers, cybersecurity leaders and state authorities to further assess ML4’s capabilities and behavior in real-world environments” — a practice that’s becoming commonplace for proprietary American labs like OpenAI, Google, and Anthropic as concerns about model capabilities mount.
Similarly to Anthropic’s Project Glasswing and OpenAI’s rollout of Astra, initial testing partners will get access to a less guardrailed version of ML4 with “expanded cybersecurity capabilities.”
Outside security, Mistral said Le Chonk excels in finance and multimodal use cases. The company said it is still waiting on final benchmarks. However, early third-party analysis shows ML4 competing on par with pricier proprietary models like GPT-6 Astra in certain computer vision tasks (like in the benchmarks below), as well as impressive open-weight Chinese models like Kimi K3. Le Chonk met or slightly outperformed DeepSeek models on financial work tasks, and hit a new high of 15% for open-weight models on Harvey’s Legal Agent benchmark.
Vals.ai via Mistral
As a reminder, benchmark scores themselves should be taken with a grain of salt, especially considering how many models cheat.
Also: The AI models that cheat the most, according to new CAIS benchmark
Chinese labs like DeepSeek and Moonshot (which develops Kimi models) have been accused of distilling, or ripping off, proprietary models from American labs to gain their competitive edge. Mistral reiterated it’s not participating in that process.
“We are fully separate from other models, and we don’t take inspiration from them,” Stock said in the briefing.
The company also leaned on its commitment to sovereignty, an equally hot topic, especially in Europe.
“Customers will soon have flexible deployment options: self-deploy or access itn sovereign region where data stays under EU jurisdiction,” Mistral wrote
More for less compute
Training a competitive model in a compute shortage is no small task for a trimmer lab like Mistral, which doesn’t have the same resources as a pre-IPO giant like Anthropic.
“ML4 was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months, deployed in Mistral’s own data centers in Europe,” the company said, adding that the preview will also run on those same GPUs. For context, Nvidia CEO Jensen Huang said on X that OpenAI trained GPT-6 Astra on roughly 100,000 GPUs. That’s quite the fraction.
“We expect the model to improve significantly over the next few months. This model will also serve as the base for a new wave of specialized and optimized models from Mistral,” the company added.
