Reuters reports Chinese military-linked researchers built defence chatbot on Meta's Llama

Reuters reported that six researchers from three Chinese institutions, including two affiliated with the People’s Liberation Army’s Academy of Military Science, had described in a June 2024 paper building an intelligence-analysis chatbot called “ChatBIT” on top of an early, openly released version of Meta’s Llama.

The researchers fine-tuned Llama 2’s 13-billion-parameter model, released by Meta in February 2023, adding their own parameters to build a tool intended to gather and process intelligence and support military operational decision-making. The paper said ChatBIT, tuned for dialogue and question-answering in the military domain, outperformed several other AI models and reached roughly 90% of the capability of OpenAI’s GPT-4 on the metrics the researchers used. Meta said it had not been aware of the project and stated that “any use of our models by the People’s Liberation Army is unauthorized and contrary to our acceptable use policy,” which bars military applications; the company did not describe any way to prevent such downstream use once weights were public. Reuters noted the tool’s training set was comparatively small, around 100,000 military dialogue records, against the trillions of tokens used to train contemporary frontier models.

The report became a frequently cited example in a live policy argument about open-weight models: once weights are published, a developer’s acceptable-use policy is a statement of intent rather than a technical control, and a state actor barred from a model’s terms of service can still fine-tune and deploy it. The case fed directly into the debate over whether open-weight releases from major US labs warranted export-style restrictions similar to those applied to advanced chips, an argument that intensified through 2024 and 2025 as Chinese labs also began releasing their own competitive open-weight models.