What happened
Mistral AI released the open weights for its 1-trillion-parameter Mistral Large 4 model, also known as Le Chonk, making it free to download, modify, and run outside the developer's control. The company acknowledged that while the model refused black-box attacks in live preview traffic, it sees a potential for uplift to conduct real attacks. Mistral stated it is modifying guardrails for its API and hosted services but did not specify how these updates would apply to weights already downloaded by users. The release includes a modified MIT license permitting commercialization, with API pricing set at $0.70 per million input tokens and $2.10 per million output tokens.
Mistral AI released the open weights for Mistral Large 4, a 1-trillion-parameter , on Thursday. The company stated that while the model refused to comply with black-box attacks against live targets during its public preview, it identified a potential for uplift to conduct real attacks. Mistral is modifying the model’s guardrails and plans to keep updating them in the coming months, but these updates will only reach users of Mistral’s API and hosted services. The company did not clarify how these guardrail updates would apply to weights already downloaded by users, who can run and improve the model on their own infrastructure without depending on Mistral.
The model is released under a modified MIT license that permits use, modification, fine-tuning, and commercialization. For organizations preferring not to host the model, Mistral’s API is priced at $0.70 per million input tokens and $2.10 per million output tokens. Mistral also announced that design partners, such as Orange Cyberdefense, are getting preview access to a specialized cybersecurity variant optimized for penetration testing and red-teaming, subject to rigorous vetting and contractual safeguards.
Mistral reported that Large 4 refused 96.6% of malicious prompts in the cyber-related portions of the JailbreakBench, StrongREJECT, and AgentHarm benchmarks, which the company claims is the highest rate measured among open models. The model scored 51% on the Artificial Analysis Cyber Index, though two other open-weight models, DeepSeek V4.1 Flash and MiMo-V2.6-Pro, scored higher. Mistral also claims Large 4 is the best open-weight model for coding, citing leading scores on Cognition’s FrontierCode 2.0 and Terminal-Bench 4.0, but has not yet published the technical report behind these results.
Why it matters
The release of a highly capable open-weight model with acknowledged cyberattack potential creates a persistent and irreversible risk of misuse, as noted by the UK’s AI Security Institute. Unlike closed models, open weights shift responsibility for guardrails, access controls, and misuse prevention entirely to the operator, complicating enterprise security strategies. While Mistral claims the model refused 96.6% of malicious prompts in specific benchmarks, analysts warn this figure does not guarantee safety in production environments against motivated attackers. The divergence between hosted API safety and self-hosted risk highlights a critical gap in AI governance, where sovereignty and security must be evaluated together rather than traded against each other.
The UK’s AI Security Institute previously reported that leading open-weight models trailed closed frontier models on cyber tasks by four to seven months, warning that open release of highly cyber-capable models creates a persistent and irreversible risk of misuse. This release underscores the tension between open innovation and security, as attackers are already using open-source AI tooling, such as the ARTEX framework, to target thousands of tenants.
Analysts caution that Mistral’s 96.6% refusal rate in benchmarks does not guarantee safety in production environments. Niyati Daftary of Gartner noted that the figure reflects performance in specific test scenarios, not behavior with diverse data or adversarial use cases. Charlie Dai of Forrester added that without a detailed technical report, the claim of being the best open-weight model for coding is difficult to validate, and leadership in this space can shift within weeks.
For enterprises, self-hosting Large 4 shifts responsibility for patching, monitoring, evaluation, and misuse prevention to the operator. Daftary advised that enterprises should run the model as a controlled component rather than an autonomous agent, using isolated environments, least-privilege access, and human sign-off before remediation. This highlights the need to evaluate sovereignty and security together, as open weights and self-hosting complicate traditional security postures.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').Why can ethical evaluation not be reduced to one model score?
What to watch next
Monitor how enterprises implement security controls for self-hosted Large 4, particularly regarding isolated environments and human sign-off for remediation. Watch for the publication of the technical report behind Mistral's coding and cyber capability claims, which is currently pending. Additionally, observe the EU’s AI Act compliance status for Large 4, as the company has not disclosed whether the model falls under the systemic risk category requiring specific risk assessments and incident reporting.
The publication of the technical report behind Mistral’s coding and cyber capability claims is pending. Until this report is available, independent validation of the model’s performance and safety claims remains limited.
Enterprise adoption patterns for self-hosted Large 4 will reveal how organizations balance data sovereignty with security risks. Watch for best practices emerging around isolated environments, sandboxed code generation, and controls for vulnerability triage.
The EU’s AI Act compliance status for Large 4 is unclear. The company has not disclosed whether the model falls under the systemic risk category for models trained with more than 10^25 FLOPs, which would require specific risk assessments, incident reporting, and cybersecurity protections.