Researchers linked to China's military used Llama 2 to develop an AI model for potential military use; Meta says use of “outdated” Llama 2 would be “irrelevant”
The report tests the practical boundary of Meta’s open-model strategy: the company had presented Llama 3.1 as a frontier-level open model only months earlier, while arguing that the older Llama 2 used here no longer mattered technologically.
It also complicates the distinction between commercial openness and state use. Related coverage shows Meta later made Llama available for US national-security applications, making access and downstream deployment—not simply whether a model is open—the central policy fault line.
First-order effects
Chinese institutions linked to the PLA can point to a publicly available Meta model as an input to a system with potential military applications; Meta is immediately pressed to defend why Llama 2’s age limits the significance of that use.
Meta’s response narrows the issue to model capability and versioning rather than disputing that released weights can be adapted by actors it does not control.
Second-order effects
The episode increases scrutiny of whether open-model licensing, release practices, and downstream-use restrictions can meaningfully constrain military adoption after weights are available.
It gives governments and enterprise buyers a concrete case for separating access to current frontier models from access to older, capable open releases; Meta’s national-security Llama partnerships make that distinction more politically salient.
Third-order effects
If similar cases recur, AI policy is likely to focus less on a binary open-versus-closed debate and more on differentiated model access, provenance, and deployment controls by user and jurisdiction.
The case suggests model generations may age faster than their strategic usefulness: an older released model can remain a reusable base for specialized systems, sustaining pressure for model-access balkanization.
The trend: Open-weight AI is becoming a geopolitical access question, as model developers try to reconcile broad distribution with national-security uses they cannot fully direct.
Another important point that this story unfortunately omits. It's not just the speed with which Chinese tech companies are releasing their own open AI models, but that those Chinese models are competitive, in terms of performance, with US models.
@SunnyCheungky @JamestownTweets In a June academic paper, researchers incl those under the PLA's leading research body, the Academy of Military Science (AMS), detailed how they had used @Meta's Llama as a base for its own “ChatBIT” AI model for battlefield information processing …
And the assertion that the model is 90% as capable as OpenAI's ChaptGPT-4. 90% as capable at doing what?! Reuters doesn't say. Does the underlying research? We have absolutely no idea.
The alleged role of a single and outdated version of an American open-source model is irrelevant when we know China is already investing over 1T to surpass the US technologically, and Chinese tech companies are releasing their own open AI models as fast, or faster, than US ones.
🇨🇳The PLA used an open source LLM to develop a chat bot. Incredible, right? This is “exclusive” journalism. 🤣 If they're SUPER thorough, maybe Reuters will discover Github next. 🤫
The PLA researchers say ChatBIT could in future be applied beyond intelligence analysis to strategic planning, simulation training and command decision-making. @Meta says such use by the #PLA is unauthorized, and a violation of its terms of use policy
Exclusive: Days before the U.S. election we show Chinese PLA-linked researchers have developed a powerful AI model for military use on back of Meta's open-source Llama w/ @JessiePang0125 https://www.reuters.com/...
This is the 1st time there's been “substantial evidence” PLA military experts in China have been systematically researching and trying to leverage the power of open-source LLMs, said @SunnyCheungky of @JamestownTweets Foundation who researches China's dual use tech including AI
Here's what's important to know about this story: many claims included in the piece about this new Chinese model don't make sense on their face, including that this new model was developed using 100k dialogues, substantially less than what other models are fine-tuned on.