Europe's Strongest Model Debuts on Cyber-Attack Capability
Europe's strongest 'sovereign AI' card happens to be the world's first trillion-parameter model to publicly sell cyber-attack capability as its headline feature. Open weights plus a low-guardrail government edition — in the name of 'digital sovereignty', Europe is issuing the world a double-edged sword.
- 01THE LAUNCH — On October 6, the public preview went live, API first, nicknamed Le Chonk.
- 02THE BENCHMARKS — Cybench 93%, CyberGym 82%: the world's highest cybersecurity benchmark scores.
- 03THE TWO TIERS — Two tiers: governments and security partners get the low-guardrail edition; the public edition carries more guardrails.
WHAT HAPPENEDWhat Happened
On October 6, 2026, France's Mistral AI released Mistral Large 4 in public preview, officially nicknamed 'Le Chonk'. The specs: a 1.05 trillion-parameter MoE with 49 billion active per token, natively multimodal, trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centres — the training infrastructure entirely European.
The biggest selling point is not parameters. It is cybersecurity: 93% on Cybench, a global-best 82% on the CyberGym-E2E vulnerability-reproduction-and-patching task, and confirmation from Artificial Analysis as the number one open-weight model outside China. Mistral itself acknowledges that several closed frontier models score close to zero on such tasks — because they simply refuse to answer offensive security prompts.
THE PITCHSelling cyber-attack capability: a first in history
No mainstream model vendor has ever dared to put 'offensive cybersecurity capability' centre stage at a launch. Mistral not only did; it gave numbers: Cybench 93%, CyberGym-E2E 82%, a global best, ahead of MiMo-V2.6-Pro and GPT-6 Luna. Agentic coding on DeepSWE at 61.7% beats DeepSeek V4 Pro; on the Harvey legal benchmark, 15% beats GPT-6 Astra.
The subtext of this play is blunt: while US models add guardrails and Chinese models compete on Chinese-language scenarios, Europe chose a differentiation track nobody dared touch — 'we do not refuse to answer dangerous questions'. In the politically correct AI-safety climate of 2026, that is heresy — and precise positioning.
THE TWO TIERSTwo tiers: fewer guardrails in the government edition
One detail of the preview deserves attention: vetted cybersecurity partners and national institutions receive a version with 'reduced moderation and expanded cyber capability' — in Reuters' wording, 'fewer safety guardrails' — while the public version carries more. Mistral states plainly that it is privately red-teaming with cybersecurity partners.
In other words: the same model, with the government unlocking full attack capability while the public gets a neutered edition. This is the ultimate form of sovereign AI — the state customises not just the model, but the model's 'danger level'. When France hands Le Chonk to an ally's cyber forces, whether it is handing over a weapon or a product depends on who is looking.
THE OPEN-WEIGHT FUTUREOpen-weight futures: the weights ship at the end of October
The API is live, but the weights are only promised for the end of October, and the licence is still unpublished. As aidailypost put it: 'API is live now... but the weights do not ship until end of October, so self hosting is not yet possible.' For now, 'open source' is a futures contract.
The pricing is calculated too: preview pricing of $1.36 per million input tokens and $4.18 per million output tokens, half price for the first two weeks. This is not selling an API; it is seizing ecosystem position — locking developers into Mistral Cloud before the weights open. And the hardware bar for self-hosting the full 1.05 trillion weights is something Mistral has still not mentioned: once truly open, how many organisations could actually run it?
THE SHIFTWhat Changed
SIGNAL JUDGMENT — ANALYSISIn sovereignty's name, a double-edged reality
When France writes 'we do not refuse dangerous questions' into its national AI strategy, it is betting that the geopolitical leverage of offensive capability outweighs the accountability of proliferation risk. History repeatedly shows that once a weapon is open-sourced, the first to lose control of it is its inventor. Le Chonk's bulk was fed by a hundred kitchens; what is left after the guardrails come off is its lean side.
WHAT REMAINS UNCERTAINWhat Remains Uncertain
EVIDENCEEvidence
The public preview went live on October 6; Cybench 93% and CyberGym-E2E 82% are officially published figures, and Artificial Analysis independently confirmed it as the top open-weight model outside China; the weights release is a commitment (end of October), with the licence undecided.
PRIMARY SOURCESPrimary Sources
- Artificial Analysis: Mistral Large 4 France AI
- Reuters: Weights to open on October 27 (as reported)
- thehype.news: Mistral Large 4 benchmarks
- artificialanalysis.ai: Original article
https://artificialanalysis.ai/articles/mistral-large-4-france-ai
RELATEDRelated Intelligence
- Microsoft Winds Down Its Oracle AI Cloud Rental as Older Nvidia Chips Fall Behind
- An Unreleased Anthropic Model Proved an 18-Year-Old Quantum Conjecture, Machine-Checked in Lean
- IBM Unveils Quantum System Two in Chicago, Its Most Advanced Computer Yet, Powered by Nighthawk
- Vistra Posts $1.14B Q3 Profit as Power EBITDA Jumps Over 50%, Lifts 2026 Outlook