OpenAI’s Open-Weight Models Set to Join the US Military Ranks

When OpenAI revealed its first open-weight models in years this August, attention wasn’t limited to tech firms. The release also garnered interest from US military and defense contractors, who viewed it as an opportunity for highly secure operations.

Initial feedback indicates that OpenAI’s tools fall short compared to competitors in key capabilities, according to some military vendors speaking with WIRED. Nonetheless, they are pleased that a significant industry leader’s models are now available to them.

Lilt, an AI translation company, collaborates with the US military to process foreign intelligence. Due to the sensitive nature of the information handled by its software, it must be installed on government servers and function without an internet connection, a practice referred to as air-gapping. Previously, Lilt created its own AI models or utilized open-source alternatives like Meta’s Llama and Google’s Gemma. OpenAI’s tools were previously inaccessible due to their closed-source nature and requirement for online access.

The introduction of OpenAI’s new open-weight models, gpt-oss-120b and gpt-oss-20b, has changed this situation. Both can operate locally, giving users the ability to install them on their own devices without cloud connectivity. Additionally, with access to the models’ weights—crucial parameters influencing their responses to various prompts—users can customize them for specific applications.

OpenAI’s return to the open-source arena could enhance competition and result in improved systems for militaries, healthcare providers, and others managing sensitive data. A recent McKinsey survey of about 700 business leaders found that over 50 percent reported their organizations are utilizing open-source AI technologies. Models exhibit varying strengths based on their training, and organizations often employ multiple models, including open-weight options, to ensure reliability across diverse scenarios.

Doug Matty, chief digital and AI officer for the so-called Department of War, the term the Trump administration has adopted for the Department of Defense, tells WIRED that the Pentagon is aiming to incorporate generative AI into both battlefield systems and administrative operations like auditing. Some applications will necessitate models that aren’t dependent on the cloud, he explains. “Our capabilities must be adaptable and flexible,” Matty states.

OpenAI did not reply to inquiries regarding how its open-source models might be utilized by the defense sector. Last year, the company lifted a broad ban on its technology for military and warfare applications, a decision that drew criticism from activists concerned about the potential harms of AI.

For OpenAI, providing a free and open model could bring multiple advantages. Enhanced accessibility may foster a larger community of experts skilled in its technologies. Additionally, since users aren’t required to register as formal clients, they might operate with greater confidentiality, potentially shielding OpenAI from scrutiny related to controversial clients—such as the military.