Key Points
- OpenAI released gpt‑oss‑120b and gpt‑oss‑20b, which can run locally and be fine‑tuned.
- U.S. military seeks offline AI solutions for secure data handling and battlefield integration.
- Lilt and EdgeRunner AI are testing the models for translation and virtual assistant applications.
- The Pentagon signed up to $200 million per year deals with OpenAI, xAI, Anthropic and Google for AI prototypes.
- Ask Sage provides the military access to about 125 open‑source and 25 closed AI models.
- Open models offer customizability and privacy but may hallucinate more and require expensive hardware.
- Experts debate the trade‑offs between open‑source flexibility and the performance of major cloud providers.
The US Department of Defense seal in the Pentagon Briefing Room in Arlington Virginia.
OpenAI’s Open-Weight Models
OpenAI introduced two open‑weight models, gpt‑oss‑120b and gpt‑oss‑20b, that can be installed on local hardware. By providing the model weights, OpenAI allows users to run the models without an internet connection and to fine‑tune them for specific tasks.
Military Interest and Use Cases
The U.S. Department of Defense is exploring the models for both battlefield systems and back‑office functions that require offline operation. Lilt, a translation firm that works with the military, can now consider OpenAI’s models alongside existing open‑source options such as Meta’s Llama and Google’s Gemma. EdgeRunner AI is modifying gpt‑oss for a virtual personal assistant that will be tested by the Army and Air Force.
Strategic Partnerships
The Pentagon’s “Department of War” unit has signed one‑year agreements worth up to $200 million each with OpenAI, xAI, Anthropic and Google to develop AI prototypes. Through the Ask Sage platform, the military accesses roughly 125 open‑source models and 25 closed options.
Advantages and Drawbacks
Open models offer greater control, customizability and privacy, which are valuable for sensitive government work. However, experts note that these models can hallucinate more often than leading commercial offerings and may require costly infrastructure to run at scale.
Industry Perspectives
Some analysts argue that large cloud providers such as Microsoft, Amazon and Google deliver more capable models for government tasks, while others stress the importance of independence from big‑tech suppliers. Companies like Moonshine provide perpetual copies of their models for a one‑time fee, and RAND highlights the utility of open models for precise translation in influence operations.
Source: wired.com