Knowing It Could Help Attackers, Mistral Released a Trillion Weights Anyway
On October 29, 2026, Mistral released the open weights of Large 4, nicknamed Le Chonk, with about 1.05 trillion total parameters. The company also warned the model could uplift real attack capability, and that guardrail updates will not automatically cover downloaded copies.
- 01THE PREVIEW — On October 6, 2026, Large 4 entered public preview.
- 02THE IMPROVEMENT — Three weeks of additional reinforcement learning produced the improved weights.
- 03THE RELEASE — On October 29, 2026, the weights opened for public download.
WHAT HAPPENEDWhat Happened
What was released is an improved version, not a re-upload of the October 6 preview. Mistral says three additional weeks of reinforcement learning improved coding and cybersecurity performance; the model has about 1.05 trillion total parameters, activates about 52 billion per task, has a context of 1 million tokens, and offers low, medium and high reasoning modes.
The guardrail terms are written into the release announcement. Mistral says the model refused 96.6% of requests in malicious cyber request testing, but monitoring of real traffic during the preview still showed potential uplift for real attacks. The company says it is revising the guardrails and will keep updating them in the coming months.
The key limitation sits in the weights themselves. Subsequent guardrail updates will apply only to the hosted API and future versions; copies already downloaded will not update automatically. A variant dedicated to penetration testing and red-team exercises is offered only to vetted design partners, such as Orange Cyberdefense.
CONTROL LEAVES WITH WEIGHTSOpen Weights Mean Handing Over Control
In the API era, vendors can change guardrails at any time and take models offline at any time. Once weights are public, anyone can run, fine-tune and strip restrictions on their own servers, and the vendor's later updates cannot catch up with distributed copies.
A Mistral executive's reasoning is that defenders need tools that cannot be cut off. Losing the model mid-way through a real attack is itself a risk. The conflict between openness and safety has no solution here that satisfies both sides, only trade-offs.
READ THE 96.6 PERCENT RIGHTRead the 96.6% Refusal Rate Accurately
That figure is the average across the cyber-related portions of three public test sets, JailbreakBench, StrongREJECT and AgentHarm, not all malicious requests. The Artificial Analysis cyber index score of 51 measures defensive work, not attack capability.
The preview monitoring results deserve more weight. The model refused black-box attacks against real targets, yet the company still acknowledges the possibility of uplift for real attack capability. A refusal rate and uplift can both hold at once, because they measure different links in the chain.
FREE TO DOWNLOAD, COSTLY TO RUNFree to Download, Expensive to Run
The FP8 weights download runs to multiple terabytes, and running the model requires 8 GPUs. Ordinary developers can download it, but that does not mean they can run it; the model's real users remain institutional customers.
The license has boundaries too. The MEOW license allows free use, modification and commercialization, but reselling the model as an online service requires a separate commercial license. What is free is the weights, not all commercial rights.
THE SHIFTWhat Changed
SIGNAL JUDGMENT — ANALYSISThe Moment the Weights Are Released, the Guardrails No Longer Belong to the Vendor
Tokenta Signal's judgment is unchanged: Mistral's choice hands control of a defensive tool to its users; the benefit is that defenders will not be cut off, and the cost is that the vendor can never take back the models already distributed.
WHAT REMAINS UNCERTAINWhat Remains Uncertain
EVIDENCEEvidence
Key points confirmed from the original report: on October 29, 2026, Mistral released the Large 4 weights, with about 1.05 trillion total parameters and a 96.6% refusal rate, while warning of potential uplift for real attacks; guardrail updates do not cover downloaded copies.
PRIMARY SOURCESPrimary Sources
- CSO Online: Mistral open-sources AI model despite warning it could aid cyberattacks
https://www.csoonline.com/article/4242648/mistral-releases-large-4-weights-despite-warning-of-cyberattack-uplift.html - Dealroom: Mistral releases the weights of Le Chonk
https://dealroom.co/news/168495-mistral-releases-the-weights-of-le-chonk-its-trillion-parameter-open-mod/
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