Captured before automated identity checks (Oct 14, 2026): accurate as of its capture on Jul 6, 2026; our capture on Nov 17, 2026 found the same text, but its identity was not checked by them.
Get the weekly coverage letter
Companies change their terms without announcing it. We capture governance documents, compare each new version with the last, and publish what changed, with the source text behind it. One email a week on what we found and what we could not reach.
This page describes what the document states, permits, or reserves. It does not constitute a legal determination about enforceability. Regulatory applicability may vary by jurisdiction. Methodology
This is the official model card for Meta's Llama 4 AI model collection, covering two models, Llama 4 Scout and Llama 4 Maverick, released April 5, 2025, under a custom commercial license. The card discloses that training data includes publicly shared Instagram and Facebook posts and user interactions with Meta AI, and that developers who deploy these models are responsible for performing their own safety testing and tuning for their specific applications. The card also states that use beyond the 12 explicitly supported languages or beyond five input images for visual tasks is considered out-of-scope, with developers assuming responsibility for safety in those extended use cases.
This model card governs the release and use of Meta's Llama 4 collection, comprising Llama 4 Scout (17Bx16E) and Llama 4 Maverick (17Bx128E), multimodal large language models released under the Llama 4 Community License Agreement. The card states that training data includes 'publicly available, licensed data and information from Meta's products and services,' explicitly including 'publicly shared posts from Instagram and Facebook and people's interactions with Meta AI,' and establishes that developers bear responsibility for safety testing, acceptable use compliance, and ensuring deployments in unsupported languages or extended image inputs are 'mitigated for risks.' The document asserts that use in languages or capabilities beyond the 12 explicitly supported is out-of-scope unless developers independently ensure safe and responsible use, and that all reported evaluations were conducted on bf16 models while quantized checkpoints are provided for deployment, creating a potential gap between evaluated and deployed model performance. The card engages AI governance frameworks including the EU AI Act, and its disclosure of training data sourced from Meta platform user interactions may require evaluation under GDPR and CCPA in the context of downstream commercial deployments; the card's open-source release structure means training compute and emissions costs are borne by Meta and not replicated by deployers. The model card's safety obligations are structured as developer responsibilities rather than platform-enforced constraints, meaning compliance with the Acceptable Use Policy and safe deployment practices depends on downstream implementers rather than technical enforcement at the model level.
The model card discloses that Llama 4 training data includes publicly shared posts from Instagram and Facebook and people's interactions with Meta AI, which may be relevant to individuals who have used those platforms or interacted with Meta AI. The card establishes that developers deploying Llama 4 in applications bear primary responsibility for safety testing and acceptable use compliance, meaning the safety profile of any given Llama 4-powered product depends substantially on the implementing developer's practices. The document does not specify a consumer-facing opt-out mechanism for the use of platform interaction data in model training, but references Meta's Privacy Center for additional information at facebook.com/privacy/guide/genai/.
Which mapped governance frameworks each document engages, tied to the specific provisions that engage them.
We withheld clauses we had published from this document. Each withholding is recorded in our correction log:
“before deploying any applications of Llama 4 models, developers should perform safety testing and tuning tailored to their specific applications of the model.”
“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 …”
“A mix of publicly available, licensed data and information from Meta's products and services. This includes publicly shared posts from Instagram and Facebook and people's interactions with Meta AI.”
“We conduct recurring red teaming exercises with the goal of discovering risks via adversarial prompting and we use the learnings to improve our benchmarks and safety tuning datasets. We partner early with subject-matter experts in critical risk areas to understand …”
“The Llama 4 model collection also supports the ability to leverage the outputs of its models to improve other models including synthetic data generation and distillation. The Llama 4 Community License allows for these use cases.”
We have not recorded a change to this document since we began capturing it. Monitor includes alerts from a daily check, structured change summaries, and monitoring for up to 20 platforms.
Need provision-level monitoring and regulatory mapping? Insight includes governance timelines, drift analysis, and full provision tracking.
Cross-platform context
See how other platforms handle Developer Responsibility for Safety Testing and similar clauses.
Compare across platforms →ConductAtlas detected a major restructuring of Meta’s privacy policy that removed detailed consumer rights disclosures and relocated them t…
Governance Monitoring
Structured alerts for policy changes, governance events, and provision updates across 343 platforms.