Tech • AI • Robotics • Game
Le Chonk: Mistral launches its big kitten
8/10TL;DR
Mistral has unveiled Mistral Large 4, a new European flagship AI model designed for enterprise use, while separate controversies over stolen-data marketplaces, AI privacy, model cheating and autonomous weapons highlight the risks surrounding the sector.
KEY POINTS
Mistral returns with a flagship model
Mistral Large 4 is the company’s first new frontier model in nearly a year and marks a bid to re-enter direct competition with leading AI labs. Internally nicknamed the “Chonk,” it is a mixture-of-experts system with 1 trillion parameters in total but only 49 billion active parameters per query, a design meant to cut inference costs while preserving high performance.
Built for enterprise workloads
The model accepts text and images as input and produces text outputs, with support for reasoning and code generation. It offers a 1 million-token context window and is positioned especially for large professional tasks in finance, industry, and cybersecurity, where it is presented as particularly competitive for European corporate users.
Sovereignty is central to the pitch
A key selling point is infrastructure control: the model is hosted in Europe on Mistral data centers, and the company says it was trained entirely in France on its own infrastructure. That positioning is aimed at organizations worried about dependence on American or Chinese providers, especially in regulated sectors such as banking, manufacturing and aerospace.
Open weights, but not yet fully released
Mistral Large 4 is already available by API and in the company’s developer playground, while open-weight release is scheduled for the end of October. The exact commercial licensing terms remain unclear, and local deployment will be limited to well-equipped organizations because running it requires roughly 500 GB of memory, or about four to five NVIDIA B200 GPUs.
Aggressive pricing targets adoption
The launch includes discounted API pricing at about $0.68 per million input tokens and just over $4 per million output tokens, with list prices later rising to around $1.36 for input. That makes the model notably cheaper than many top-tier rivals, even if it does not claim leadership on every general-purpose benchmark.
Competitive, but not dominant
On broad comparisons, the model appears to rank among the leading Western open-weight systems, though the most recent top Chinese open models still set a higher bar on some capabilities. Its significance lies less in outright global leadership than in proving that a European lab can remain in the race with far fewer compute resources than its largest competitors.
A stolen-data marketplace draws official scrutiny
Separately, attention has turned to Breach Forums-style services selling access to leaked personal data. One site, BrickHub, was described as indexing more than 14 billion records compiled from earlier hacks and data leaks, including addresses, family links, emails and other sensitive information. French authorities have been alerted, but the site remained accessible at the time of reporting.
Removal offers are viewed as extortion
The platform reportedly offers paid “deletion” services for victims who want their records removed. Cybersecurity observers warn that such payments are effectively worthless because the underlying data has already circulated elsewhere, meaning a copy can reappear on other services even if one listing is deleted.
Claude case raises privacy questions
A separate case involving Anthropic’s Claude has renewed concern over how “private” AI conversations really are. A Florida woman who used the chatbot like a personal diary was arrested after messages interpreted as threats were escalated internally and then reported to law enforcement. The case underlines that opting out of training does not prevent safety systems or human reviewers from accessing flagged exchanges.
Ukraine deploys AI-guided defensive guns
In warfare, Ukraine has begun deploying AI-guided heavy machine-gun turrets to intercept fast Geran-5 attack drones, which can reportedly travel at around 600 km/h. The systems automate targeting and fire control to hit threats too fast for human operators, illustrating how AI is moving deeper into battlefield decision loops even when framed as defensive technology.
CONCLUSION
Mistral Large 4 gives Europe a credible new high-end AI model and a stronger claim to technological autonomy. At the same time, the week’s parallel stories show that performance gains in AI are arriving alongside mounting concerns over privacy, cybercrime and automated force.
Ask a question