OpenAI’s Open-Weight Models Are Coming to the US Military (www.wired.com)

OpenAI’s newly released open-weight models, gpt-oss-120b and gpt-oss-20b, can now be run locally with access to model weights, making them appealing to the US military and defense contractors that require air-gapped, customizable AI for classified or offline operations. Early adopters like Lilt, EdgeRunner and Vector 35 have begun testing or integrating the models for translation, document processing and assistant tools; the Pentagon has also signed prototype deals with multiple AI vendors to explore battlefield and back-office use. OpenAI’s pivot back into open-weight releases — after reversing its previous ban on military use — gives defense users freedom to fine-tune, host privately, and combine models to meet niche operational needs. Technical trade-offs matter: gpt-oss variants are text-only, sometimes underperform on certain languages and low-compute environments, and currently lag more capable cloud-hosted commercial models on robustness and multimodal tasks. Adopters mitigate this by caching domain data or chaining models, but critics warn open models can hallucinate more and require costly infrastructure to run at scale. Supporters counter that open weights deliver critical control, privacy, and supplier independence for edge scenarios (drones, satellites, classified networks). The release intensifies competition in open-source AI, forcing choices between control/customizability and raw performance — a balance the military will weigh as it moves from demos to operational testing.