OpenAI’s return to open-weight AI has quickly become more than a technology-industry story. Its new models, gpt-oss-120b and gpt-oss-20b, are now drawing attention from US military offices and defense contractors that need AI systems capable of working in highly controlled environments.
The appeal is straightforward: these models can run locally. That means they can be installed on users’ own devices or servers without requiring a cloud connection, and their weights can be adjusted for specific tasks. For organizations handling sensitive government information, that changes where OpenAI can fit.
Why local AI matters for defense
Some military work cannot depend on an online AI service. Lilt, an AI translation company that contracts with the US military to analyze foreign intelligence, must install its software on government servers and operate without an internet connection. That kind of setup is known as air-gapping.
Before OpenAI released gpt-oss-120b and gpt-oss-20b, Lilt used its own AI models or open source options such as Meta’s Llama and Google’s Gemma. OpenAI’s tools were not a fit because they were closed source and available only online.
The new open-weight models change that basic equation. They give military vendors another option for systems that need privacy, control, and offline operation. That does not mean OpenAI’s models are automatically better for every job, but it means they can now be considered for work that previously excluded them.
OpenAI’s broader move back into the open-source market could also increase competition. The source article notes that in a recent McKinsey survey of roughly 700 business leaders, more than 50 percent said their organizations use open source AI technologies. Many organizations use several models together because different models have different strengths.
The Pentagon wants adaptable AI
Doug Matty, chief digital and AI officer for the so-called Department of War, the name the Trump administration is using for the Department of Defense, tells WIRED that the Pentagon plans to integrate generative AI into battlefield systems and back-office functions like auditing.
Some of those uses will need models that are not tied to the cloud. Matty’s explanation was direct: “Our capabilities must be adaptable and flexible.”
Earlier this year, Matty’s unit at the Pentagon struck one-year deals worth up to $200 million each with OpenAI, Elon Musk’s xAI, Anthropic, and Google. The goal is to build prototypes of AI systems for different purposes, including automating war-fighting tools.
Until OpenAI’s recent launch, Google was the only new tech partner in that group offering a cutting-edge open model as an option. Other companies license cloud-run models that cannot be customized to the same extent as open models.
Early results are useful but uneven
Defense vendors are not describing gpt-oss as a finished answer to every military AI problem. In Lilt’s case, CEO Spence Green says a military analyst may ask a system to translate documents to English, avoid mistakes, and route the work to the most knowledgeable person about hypersonics.
Lilt’s proprietary models, trained for government applications, handle the translation. Google’s Gemma currently automates routing information to models, analysts, and other teams. The goal is to address a shortage of language experts and a backlog of data to process.
OpenAI’s latest open source models are not ideal for Lilt’s needs. They process only text, while the military also needs to sort through images and audio. Lilt also found that the models underperform in some languages and in situations with limited computing power.
Green still sees value in the added competition. “With gpt-oss, there’s a lot of model competition right now,” Green says. “More options, the better.”
Other military suppliers report more positive early results, though they are still describing an early stage. Jordan Wiens, cofounder of Vector 35, which supplies reverse engineering tools to the Pentagon and has integrated gpt-oss into its offerings, says, “It’s pretty early.”
EdgeRunner AI, which is developing a virtual personal assistant for the military that does not require a cloud connection, says it achieved sufficient performance with gpt-oss after feeding it a cache of military documents to modify its capabilities, according to a paper the company published in October. The US Army and the Air Force will begin testing the modified model this month, says Tyler Saltsman, EdgeRunner’s CEO.
Open models bring control and tradeoffs
Open models may be especially useful where fast responses or internet interference are concerns. That includes AI systems running on drones or satellites, says Kyle Miller, a research analyst at Georgetown University’s Center for Security and Emerging Technology.
Miller says open source AI models offer the military “a degree of accessibility, control, customizability, and privacy that is simply not available with closed models.”
There is also a broader ecosystem beyond direct deals with AI providers. Nicolas Chaillan, founder of Ask Sage and a former chief software officer for the US Air Force and Space Force, says the military has access to about 125 open source models and about 25 closed options through Ask Sage.
Chaillan argues that open source models have serious drawbacks for the US military. He says they hallucinate and make incorrect predictions more often than the best commercial models. He also says the infrastructure required to run the largest models can cost the same or more than licensing a commercial model over the cloud.
“It’s like going from PhD level to a monkey,” Chaillan says. “If you spend more money and get a worse model, it makes no sense.”
His view is that the military should monitor open models but focus on more capable options from Microsoft, Amazon, and Google through cloud networks developed specifically for sensitive government tasks.
The strategic question is dependence
Other military suppliers and experts disagree with a cloud-first approach. Their concern is that closed models can create dependence issues and may not satisfy the military’s specialized needs.
Pete Warden, who runs the transcription and translation technology developer Moonshine, says his contacts in the defense world have become more cautious about trusting big tech companies after seeing how Musk used his Starlink satellite network to influence government leaders.
“Independence from suppliers is key,” Warden says.
Warden’s solution is to let government agencies control a perpetual copy of Moonshine’s model in exchange for a one-time fee. That reflects the core tension around OpenAI’s gpt-oss release: military buyers want stronger models, but they also want control over how and where those models run.
For OpenAI, a free and open model may expand the community of people working with its technology. It may also let users operate without becoming formal customers, which could make some use cases less visible. OpenAI did not respond to requests for comment about how its open source models may be used by the defense industry.
The result is a still-developing market where gpt-oss is significant less because it solves every military AI need and more because it gives defense users another path. The models are now part of a larger debate over performance, privacy, cloud dependence, customization, and cost.