OpenAI's Open-Source Models Tested for Military Use, Lag Behind Competitors

In November 2025, OpenAI's recent release of its first open-weight models in years has sparked interest from the US military and defense contractors. These new models, gpt-oss-120b and gpt-oss-20b, can run locally without an internet connection, allowing them to be installed on government servers for highly secure operations.

However, some military vendors tell WIRED that OpenAI's tools currently lag behind competitors in the desired capabilities for their applications. Lilt, an AI translation company that contracts with the US military, previously used open-source options like Meta's Llama and Google's Gemma, but OpenAI's closed-source models were off the table.

Despite the initial performance gap, defense officials are still pleased that models from a key industry leader like OpenAI are now an option for them. The Pentagon plans to integrate generative AI into battlefield systems and back-office functions, some of which will require models not tied to the cloud.

Experts believe OpenAI's return to the open-source market could ultimately increase competition and lead to better performing systems for militaries, healthcare companies, and others working with sensitive data. This could benefit a wide range of organizations that often use multiple AI models, including open-weight ones, to ensure reliability across various situations.