While proprietary AI models such as OpenAI’s ChatGPT and Google Gemini remain popular, the tide may be shifting to open models as IT leaders move to customize AI and control costs.

Sometimes known as “open-weight models,” the alternatives to large language models (LLMs) can provide decision-makers with better visibility and control over internal AI use, which closed models do not, analysts said. They can also help IT leaders control the economics and governance of AI within their organizations.

“It’s almost like these blank canvases are available now and then you can paint it on your own,” said Deepak Seth, senior director analyst at Gartner. “You don’t have to make the canvas itself. So you’re not starting from scratch, even when you’re building your own model.”

Open models are free to download and use. Users can tweak and deploy them to meet their own requirements, similar to the way Linux OS is available for anyone to download, tweak, and use.

Open models have been gaining traction because more and more use cases for them are emerging, said Jesse Williams, cofounder and COO at Jozu, an AI tooling company. “Open source is more flexible and can be used in ways that proprietary models … in some cases can’t be trusted to operate,” Williams said.

“Proprietary models are gaining in usage and traction faster than any technology we have ever seen and aren’t showing any sign of slowing down,” he said, stressing that open models’ growing popularity should not be seen as simply a backlash to LLMs.

Some popular open models include Meta’s Llama, Mistral, DeepSeek and Minimax. Proprietary generative AI providers have also released open source versions of their LLMs:  Google’s Gemma is an offshoot of Gemini, OpenAI has GPT-OSS, and Microsoft has Phi.

Though the models can be fine-tuned to meet specific corporate needs, they do not share the data sources on which they were trained.

Compared to the proprietary LLMs, the open models offered by Google and OpenAI are trained on lower volumes of data, and may not be as smart, said Max Leaming, head of data science and AI solutions at ManpowerGroup. As a result, open models require experimentation to find the right application. “What is this model good at?” he said. “You have to figure that out. None of them are truly general purpose models.”

Companies that include ServiceNow, Microsoft, HubSpot and RWS have contended that open models are easier to blend into AI infrastructures, lower computing costs, and friendly to agentic AI workflows.

Recent outages at closed-model providers such as Anthropic and OpenAI are also forcing CIOs to think about vendor lock-in and the need to add open models to build AI resiliency, said Max Goss, senior research director at Gartner. “It is still early in the AI race,” he said, adding, “CIOs do need to be mindful of what they’re putting AI to work for and what is the alternative, what is the fallback plan?”

Open models are often a good option for on-premise deployment because of lowered costs and better security. “I’m also not exposing our data to the provider who may use that data to … quite literally, train models for use by our competitors,” said ManpowerGroup’s Leaming.

The models can also be used in robotics, where they provide a universal language for robots to communicate, according to Rev Lebaredian, vice president of physical AI simulation at Nvidia. He spoke at a press briefing ahead of the company’s GTC trade show earlier this year.

Nvidia’s AI robotics stack is largely open source, which is “able to connect the entire robotics ecosystem together,” Lebaredian said.

Open models are also playing a role in the pu sh for digital sovereignty in Europe and elsewhere. France, for example, is pinning its sovereign AI strategy on Mistral, while the UAE has K2 Think V2, which was developed by Mohamed bin Zayed University of Artificial Intelligence, G42, and chip company Cerebras Systems.

Open models are important to sovereign AI so nations can understand, adapt, and control systems powering digital infrastructure, said Richard Morton, vice president and managing director of the Institute of Foundation Models at MBZUAI.

MBZUAI’s K2 Think V2 gives countries the ability to build AI aligned to their own priorities, languages, values, and security needs. “For us, sovereignty is ultimately about meaningful ownership of the technology itself,” Morton said.

Despite increased interest in them, open models do carry some security risks. Bad actors could hack systems with malicious prompts, for instance, or use the AI technology to launch attacks, according to a study published by the UK Department for Science, Innovation and Technology and the AI Security Institute.

Flawed models with vulnerabilities could serve as an entry point for hackers to break into corporate systems, according to the study, which was chaired by Yoshua Bengio, a deep learning pioneer. “Unlike closed models where hosts can universally roll out fixes, open-weight model developers cannot guarantee that updates will be adopted by users,” the study said.