Mistral AI has released the open weights of Mistral Large 4, the trillion-parameter multimodal model it calls "Le Chonk," making it free to download, modify and run outside the company's control.

In its announcement, the Paris-based company warned that preview monitoring, while showing the model refusing black-box attacks against live targets, had revealed "a potential for uplift to conduct real attacks."

Mistral said it will keep updating the model's guardrails in the coming months, but vice president of science Pierre Stock told CSO that subsequent safeguard updates are applied to the managed API and future releases rather than to downloaded checkpoints — meaning already-downloaded copies will never receive the fixes.

The release, which landed weeks after the model's public preview and has drawn broad coverage from CSO, TechCrunch, ZDNET and The 420.in, frames open weights as both a capability milestone and a deliberate acceptance of risk.

It is unusual for an AI developer to release a frontier model while simultaneously publishing an admission that it could help adversaries mount real attacks.

CSO notes Mistral's caution follows a UK AI Security Institute finding that leading open-weight models trail the closed frontier on cyber tasks by four to seven months, and that for highly cyber-capable models open release creates "a persistent and irreversible risk of misuse."

Mistral's bet, as Stock told TechCrunch, is that defenders deserve frontier intelligence "that can't be cut off," arguing enterprises and governments should be able to run the model on-prem to identify and respond to incidents.

CSO also reports Microsoft recently said a China-aligned group used the agentic ARTEX framework against more than 2,250 Microsoft 365 tenants, underscoring that attackers already weaponize open AI tooling.

Mistral says Large 4 refused 96.6% of malicious prompts in the cyber portions of the JailbreakBench, StrongREJECT and AgentHench benchmarks — its highest measured rate among open models — though CSO gives the benchmark name as AgentHarm.

CSO quotes Gartner principal analyst Niyati Daftary cautioning that "Mistral's 96.6% figure tells you what their model does in their test scenario, not how it will behave with your data."

Mistral also reports a 51% score on the Artificial Analysis Cyber Index, where CSO notes two open-weight rivals, DeepSeek V4.1 Flash and MiMo-V2.6-Pro, score higher.

On coding, ZDNET reports Terminal-Bench 4.0 jumped from 28% to 41%, edging Opencode's GLM-5.3 by one point, while the overall Coding Agent Index places ML4 one point behind it at 48.

The open weights move responsibility for safety onto whoever deploys the model.

Forrester vice president Charlie Dai told CSO that self-hosting shifts ownership of guardrails, access controls, monitoring and misuse prevention to the operator, and that for European buyers sovereignty "should not outweigh security concerns."

Daftary recommends enterprises run Large 4 as a controlled component rather than an autonomous agent: isolated environments, least-privilege access to code and credentials, sandboxed generated code and human sign-off before remediation.

CSO also reports Mistral has not said whether Large 4 falls under the EU AI Act's presumption of systemic risk for models trained above 10^25 FLOPs, though Stock said Mistral is a full signatory of the EU's General-Purpose AI Code of Practice.

Mistral is pitching the release as a sovereignty play.

TechCrunch quotes Stock arguing that access "can't be cut off by definition because the model is open-weighted," and the company says ML4 was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months in its own European data centers — training scale ZDNET contrasts with the roughly 100,000 GPUs it says Jensen Huang claimed for GPT-6 Astra.

The 420.in reports Orange Cyberdefense has partnered to integrate a cybersecurity-focused variant into its products, and CSO notes vetted design partners get preview access to a specialized version for penetration testing, vulnerability assessment and red-teaming, limited to systems partners are authorized to test.

Mistral has not yet published the technical report behind its coding and cyber claims; Stock told CSO it will document the company's cyber evaluation methods, refusal testing and preview monitoring findings, but gave no publication date.

Dai told CSO that without that report, the "best open-weight model" claim remains difficult to validate across workloads, and he noted frontier leadership can shift within months or even weeks.

Meanwhile, ENISA, the EU's cybersecurity agency, said in a statement published with the announcement that it is testing the model to identify vulnerabilities, audit code and analyze defensive logs.

The modified MIT license permits use, modification, fine-tuning and commercialization, while businesses reselling the model as a service need a separate license; Mistral's API costs $0.70 per million input tokens and $2.10 per million output tokens.