US Government: Llama for Secure, Open AI

Lisa Ernst · 22.09.2025 · Technology · 5 min read

The U.S. procurement agency GSA has Meta's Llama included in a government program. This enables federal agencies to use the model under uniform framework conditions. The GSA confirms this in a press release and references the OneGov initiative for faster, safer AI adoption in federal agencies.

Quelle: Original Illustration

GSA Approval of Llama

On September 22, 2025 Reuters, citing the GSA, confirmed that U.S. federal agencies may officially use Llama. Llama is listed as an “approved AI tool” in GSA structures. The GSA press release explicitly mentions the integration into the OneGov initiative and cites OMB memos M-25-21 and M-25-22 as guardrails for use and procurement. Earlier, GSA had already approved other AI providers like OpenAI, Google and Anthropic for federal use; Llama now joins them. For security context: AWS reported in June 2025 FedRAMP “High” and DoD IL4/5 for the use of Llama and Claude in AWS GovCloud, i.e., for particularly sensitive environments. Meta itself describes the OneGov framework as cost- and data-control-friendly, since Open-Weights models can be operated locally.

Understanding Llama

Llama is a family of so-called Open-Weights models from Meta. The finished model weights are freely available, allowing developers to run them locally or in a government cloud. Source code, training data and recipes are not fully open. The Open-Source Initiative (OSI) therefore explicitly does not classify Llama as Open Source; this is important for correct classification. The NIST AI Risk Management Frameworks (AI RMF) provide a guide on how agencies should manage AI risks, independent of the model type. These are available in the NIST AI 100-1 and on the NIST Website.

Quelle: unite.ai

Llama: Open-Source AI in Code Form.

Key Drivers & Context

The approval of Llama by the GSA is driven by several motives. First, cost and control considerations: Open-Weights models enable self-hosting, improving data sovereignty and cost management. The NTIA has published a background report. Second, procurement strategy matters: OMB M-25-22 emphasizes competition, performance controls and risk management in AI procurement; open models serve as a counterweight to proprietary services and reduce vendor lock-in. Additional OMB memoranda are available. Third, security is a factor: With FedRAMP/IL4-5, the hurdle to deploy such models in sensitive government environments is lowered. At the same time, governance obligations remain high: NIST’s AI RMF requires documented risk assessments, bias testing, traceability and monitoring.

Quelle: blog.lilypadnetwork.org

AI in government everyday work: Llama as a digital assistant for more efficient processes.

Fact Check & Clarifications

Confirmed: The GSA confirms the OneGov integration of Llama; Reuters reports the approval for federal agencies on September 22, 2025. The GSA press release is available here.

Confirmed: OMB M-25-21 and M-25-22 set the framework for use and procurement of AI in federal agencies.

Confirmed: AWS GovCloud has for Llama/Claude FedRAMP “High” and DoD IL4/5, enabling use in highly sensitive environments.

Unclear: “Open Source” in the narrow OSI sense. Llama is still, according to OSI, not Open Source; Meta talks about “open source/open models,” OSI disagrees. Meta's perspective can be found here. Practically, it is Open Weights, not fully open AI, as the NTIA report shows.

False/Misleading: “With the GSA approval, Llama is automatically FedRAMP-authorized.” FedRAMP authorizes cloud services, not model families per se; relevant is the authorized operating environment (e.g., AWS GovCloud with the stated authorizations) and the agency’s risk management per the NIST AI RMF. FedRAMP information is available at fedramp.gov. A discussion appears at FedScoop and the NIST AI RMF.

Quelle: kanzlei-herfurtner.de

Llama in the server room: Open-source AI as the backbone of modern government IT.

Implications for Agencies

For agencies, the approval means that Llama can now be evaluated and deployed productively under standardized frameworks, including self-hosting options in GovCloud or on-premises. This shortens data paths and reduces vendor lock-in, as emphasized in OMB M-25-22. Agencies should review strictly along the NIST AI RMF (Govern, Map, Measure, Manage) and use the playbooks from the NIST AI Resource Center. The NIST AI RMF provides the baseline. For business units, Open-Weights enable granular cost control (GPU time, scaling), clear logging/audit trails and integrations into business processes, as long as evaluation, prompt-safety concepts and red-teaming are established. Additional information can be found in the NTIA report on Open Model Weights.

Open Questions & Future Outlook

There remain open questions. How will agencies operationalize license and compliance specifics of Llama, such as the 700-million MAU clause and Acceptable-Use Terms, especially in multi-party projects and procurements? The license explicitly sets the threshold and may require separate approval by Meta. How will buyers integrate OMB M-25-22 (AI procurement) and NIST AI RMF (risk management) into long-running framework contracts and documentation? And which evaluation benchmarks will become standard in the public sector to fairly compare Open-Weights with proprietary services, as hinted in America's AI Action Plan?

The approval of Llama for U.S. agencies marks a practical step toward greater choice: Open Weights, self-hosting and standardized procurement paths can reduce dependencies and may lower costs and data risks if governance and security architectures are implemented consistently. This is evidenced by the Reuters report, the GSA press release and OMB M-25-22. If you intend to deploy Llama, take the OSI debate seriously, review the license carefully, and use the NIST risk management as a guiding thread across project, procurement and operations. In this way, the benefits of open models in the public sector can be realized solidly and transparently.