Meta extends Llama access to France, Germany, Italy, Japan, South Korea, NATO and the EU for defense use

What the access expansion actually covers

Meta says it began making Llama available to US government agencies, including defense and national security applications, last year, and extended it to the Five Eyes partners — Australia, Canada, New Zealand, and the UK — since late last year.

The new step adds France, Germany, Italy, Japan, and South Korea, plus NATO and European Union institutions. Meta frames this as deliberate and gated: a "step-by-step approach" with further countries considered "in consultation with the US government."

So this is not a general open-source release changing hands. Llama's license already lets anyone download the weights. What changed is Meta's explicit endorsement of a specific set of allied governments for defense and national security use, and the partner ecosystem it is naming to support them.

The deployment argument is the real content

The most concrete claim in the announcement is technical, not diplomatic. Meta argues Llama fits sensitive use cases precisely because it is open weights and can be run without sending data to a third party.

Governments can also fine-tune Llama models using their own sensitive national security data, host them in secure environments at various levels of classification, and deploy models tailored for specific purposes on-device in the field.Montana Labs

That sentence describes three properties an API model cannot offer: fine-tuning on classified data that never leaves your control, hosting inside air-gapped or classified networks, and inference on hardware in the field with no network dependency.

For applied teams, this is the honest case for open weights over closed frontier APIs — not that the model is better, but that data residency and offline operation are non-negotiable in these environments. The frontier lab that wins here is the one whose weights you can physically place behind your own perimeter.

The integrator list tells you who actually ships this

Meta names its delivery partners: Accenture, AWS, AMD, Anduril, Ask Sage, Booz Allen, C3 AI, Circus, Cyberspatial, Databricks, EdgeRunner AI, Google Cloud, IBM, Microsoft, Lockheed Martin, Oracle, Palantir, Scale AI, Snowflake, and others.

That roster is the tell. Meta supplies weights; the integration, fine-tuning, hosting at classification levels, and accreditation come from cloud providers, defense primes, and data platforms. The named example — a pilot with the Army's Combined Arms Support Command using AI plus AR/VR to speed equipment repairs — is a mission-specific application, not a chatbot.

Meta also cites its Anduril partnership on wearable products for soldiers, which sits outside the Llama access story but signals how it wants the whole portfolio read: perception and decision-support hardware alongside the model.

Open weights as a lever on allied standards

Meta's closing rationale is that widespread adoption of open models like Llama will keep "our shared values" embedded in "the systems and standards adopted elsewhere," and it ties this to the US AI Action Plan it endorses and the Political Declaration on Responsible Military Use of AI.

The strategy is legible: if allied militaries standardize on Llama, they standardize on an American stack rather than a rival one, and Meta's model becomes infrastructure with policy weight behind it.

The implication for anyone building on Llama is that its governance is now entangled with US foreign policy. Access is granted country by country in coordination with the government, which means the terms under which this base model reaches sensitive users are political decisions, not just license terms — a dependency worth pricing into any long-horizon build.