Get the research letter
Companies change their terms quietly. We read every version and catch what actually changed. One email a week on the changes that matter and what they mean.
The card designates use of Llama 4 in languages beyond the 12 explicitly supported, or with more than 5 input images, as outside the evaluated scope, with developers assuming full responsibility for safety in those extended use cases.
This analysis describes what Meta's agreement states, permits, or reserves. It does not constitute a legal determination about enforceability. Regulatory applicability and practical outcomes may vary by jurisdiction, enforcement context, and individual circumstances. Read our methodology
This provision creates an explicit responsibility boundary: Meta's safety evaluations cover 12 languages and up to 5 input images, and any deployment extending beyond those parameters operates without the safety assurances described in the model card. Deploying organizations must document their own testing for extended configurations.
Under this clause, applications built on Llama 4 that operate in languages beyond the 12 supported or process more than 5 images simultaneously have not been evaluated by Meta for safety or performance, and the implementing developer is solely responsible for risk mitigation in those configurations.
Cross-platform context
See how other platforms handle Out-of-Scope Language and Image Use and similar clauses.
Compare across platforms →Monitoring
Meta has changed this document before.
Receive same-day alerts, structured change summaries, and monitoring for up to 20 platforms.
"Developers may fine-tune Llama 4 models for languages beyond the 12 supported languages provided they comply with the Llama 4 Community License and the Acceptable Use Policy. Developers are responsible for ensuring that their use of Llama 4 in additional languages is done in a safe and responsible manner. Llama 4 has been tested for image understanding up to 5 input images. If leveraging additional image understanding capabilities beyond this, Developers are responsible for ensuring that their deployments are mitigated for risks and should perform additional testing and tuning tailored to their specific applications.Excerpt from Meta's Llama 4 Model Card
(1) REGULATORY LANDSCAPE: This provision engages the EU AI Act's requirements for technical documentation and conformity assessment, which do not recognize delegation of safety evaluation to downstream deployers as sufficient for compliance in high-risk application categories. GDPR's data minimization and purpose limitation principles may also be relevant where extended language or image processing involves personal data. (2) GOVERNANCE EXPOSURE: Medium. The explicit acknowledgment that safety testing did not cover all deployment configurations creates a documented gap that deploying organizations must address in their own AI governance documentation. This is particularly material for organizations in regulated industries or those deploying in EU jurisdictions where AI Act obligations apply. (3) JURISDICTION FLAGS: EU AI Act deployer obligations are heightened for configurations not covered by the model provider's evaluations. Organizations deploying in languages not among the 12 supported should assess whether those markets have specific AI or consumer protection regulations that create additional obligations. (4) CONTRACT AND VENDOR IMPLICATIONS: Enterprise agreements incorporating Llama 4 for multilingual or high-image-volume applications should explicitly address the responsibility allocation for safety testing in extended configurations. Procurement teams should assess whether vendor representations about model capabilities align with the scope limitations disclosed in this model card. (5) COMPLIANCE CONSIDERATIONS: Compliance teams should maintain records of all deployment configurations, including languages and image input parameters, and document safety testing conducted for any configuration exceeding the evaluated scope. Legal teams should assess whether the scope limitation language affects warranty or indemnification positions in downstream customer agreements.
Regulatory citations, enforcement risk, and due diligence action items.
Provision-level monitoring, governance timelines, and regulatory mapping built from archived source documents and historical version tracking.
This provision creates an explicit responsibility boundary: Meta's safety evaluations cover 12 languages and up to 5 input images, and any deployment extending beyond those parameters operates without the safety assurances described in the model card. Deploying organizations must document their own testing for extended configurations.
Under this clause, applications built on Llama 4 that operate in languages beyond the 12 supported or process more than 5 images simultaneously have not been evaluated by Meta for safety or performance, and the implementing developer is solely responsible for risk mitigation in those configurations.
No. ConductAtlas is an independent monitoring service. We are not affiliated with, endorsed by, or sponsored by Meta.