French artificial intelligence laboratory Mistral has officially launched its latest large language model, Mistral Large 4 (ML4). Nicknamed “Le Chonk” due to its trillion-parameter scale, the model is being marketed as a premier open-weight solution designed to give enterprises and governments full sovereignty over their AI cybersecurity infrastructure.
The release comes amid growing tensions between open and proprietary AI models regarding security. Following a summer marked by significant AI security incidents, industry leaders have increasingly questioned the safety of relying on closed systems from giants like OpenAI and Anthropic. Mistral argues that open models offer a more democratic and reliable defense mechanism, allowing users to customize and own their security tools without the risk of vendors arbitrarily rescinding access.
“The cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyberattacks,” said Guillaume Lample, co-founder of Mistral. He emphasized that ML4 represents the beginning of a new generation of customizable, open-weight cybersecurity models that eliminate vendor lock-in.
Performance and Benchmarks
Mistral claims that ML4 outperforms top-tier models from competitors such as Kimi, DeepSeek, and Meta in absolute cyber capabilities. Early third-party analysis suggests the model competes favorably with proprietary offerings like GPT-6 Astra in specific computer vision tasks and matches the performance of prominent open-weight Chinese models. Additionally, Le Chonk met or slightly surpassed DeepSeek models in financial work tasks and achieved a new high of 15% on Harvey’s Legal Agent benchmark for open-weight models.
The company also addressed concerns about model origins, stating that it does not distill or derive inspiration from American proprietary models, a practice some Chinese labs have been accused of. Mistral highlighted its commitment to European data sovereignty, offering deployment options that allow data to remain under EU jurisdiction.
Efficient Training and Availability
One of the most notable aspects of ML4 is its training efficiency. Mistral trained the model from scratch using only 4,000 Nvidia Grace Blackwell GPUs over a two-month period within its own European data centers. This stands in stark contrast to reports that OpenAI utilized approximately 100,000 GPUs for its GPT-6 Astra model.
Le Chonk is currently available in public preview, with the full model weights scheduled for release on October 27. This interim period will be used for collaborative assessment with developers, cybersecurity experts, and state authorities. Similar to launches by Anthropic and OpenAI, initial testing partners will receive a version with expanded cybersecurity capabilities and fewer guardrails.
Mistral expects the model to improve significantly over the coming months and intends to use ML4 as the foundation for a new wave of specialized and optimized models.
Can’t wait for the October release. I’m tired of vendors arbitrarily rescinding access to our security data.
The GPU count difference is staggering. Four thousand versus one hundred thousand changes the whole cost equation.
Is ‘Le Chonk’ a serious codename? Feels a bit like a meme. I hope the security is as solid as the marketing.
Finally, an open-weight option that doesn’t rely on American proprietary models. Sovereignty matters.
Trillion parameters but trained in two months? That efficiency is genuinely impressive. Kudos to the team.