Open-source AI in 2025: Smaller, smarter and more collaborative

Wondering what to expect for open-source AI in 2025? Here are some developments that could shape the year ahead.

Over the past few years, AI models have been the center of attention. But in 2025, experts say the conversation will move closer to open-source AI systems.

“What we will be seeing is a shift toward AI systems, and not only models,” says Anastasia Stasenko, cofounder of pleias, a French startup that recently released Common Corpus, the largest open multilingual dataset for LLM training, as well as models trained exclusively on open data.

“What the providers will be driving value from is really the systems that are based on different kinds of open-source models,” she says. “It will be driven by the integrations, by more verticals. It's not just a model, it's a system.” Stasenko points to the elements beyond the models that feed into a larger system, like classifiers and parsers. “It's not only about the models.”

Another major conversation in 2024 was ROI for businesses investing in AI. This year, research suggests we will see a more direct correlation between ROI and open-source AI.

In an IBM study of more than 2,400 IT decision makers, 51% of businesses using open-source tools saw positive ROI, compared to just 41% of those that weren’t.

“Open-source software can give you the ability to leverage AI in a way that is cost-effective for your business and that actually gets you the velocity you're looking for,” says JJ Asghar, a Developer Advocate at IBM Research.

The Linux Foundation, known for supporting major projects like Kubernetes, Open Source Security Foundation (OpenSSF) and PyTorch, also highlighted the benefits of open source for transparency and cost-efficiency in a recent report.

“Businesses benefit from free access to powerful frameworks that eliminate high upfront costs and allow for easy customization, empowering organizations of all sizes to innovate and scale,” explains Ibrahim Haddad, a VP of Strategic Programs at the Linux Foundation.

With the passing of the EU AI Act in 2024, we can expect ethical AI to continue to be a topic of discussion and debate.

Haddad believes AI governance and regulation will play a crucial role in shaping open-source AI development. “The EU AI Act will drive ethical AI practices, pushing for transparency, fairness and accountability in open-source projects,” says Haddad. “We expect this to drive more collaboration on bias detection and ethical auditing tools.”

Stasenko says she sees “augmenting the social acceptability and the transparency for AI” as a “business opportunity in itself.” She raises the example of high-value and sensitive use cases in industries like healthcare and financial services, adding that transparency will be a cornerstone of sound risk management.

“Enterprises need a secure, safe and private way to inference with AI,” Asghar says. “The major players in the space that are not open source say they don't take your data, but enterprises can't trust that.” Asghar adds that open models like IBM Granite 3.1 can help enterprises leverage AI in a safe and private way. “This is insanely powerful,” he says.

Many companies have successfully introduced small language models in the past two years. And those models keep getting smarter and smarter. Take, for instance, Meta’s Llama models—the company recently launched Llama 3.3 70B, a text-only model that offers similar performance to the 3.1 405B model released only five months prior.

“This is also true generationally,” says Sy Choudhury, Director of AI Partnerships at Meta. “We expect these levels of advances to continue, which will benefit everyone in reducing cost and system power when delivering generative AI-based use cases to consumers and businesses alike.”

Matt White, Executive Director of the PyTorch Foundation, agrees. “I think the most pervasive trend in open-source AI for 2025 will be improving the performance of smaller models and pushing AI models to the edge,” he says, adding that smaller models can also help ensure less reliance on external systems.

“There are a lot of hard problems in the field of deep learning,” White asserts. “Hopefully, there will be innovations that bring models down in size while maintaining performance, making models more responsive and keeping a model’s knowledge up-to-date, without the reliance on external systems.”

Stasenko stresses that if 2024 was the year of the rise of smaller models, 2025 will be the year of leaning into the capabilities of every model, with a focus on energy efficiency. “This is the only way,” she says.

Multimodal models—or models that can process multiple, different types of media, such as text, videos, images and audio—will also get better and more ubiquitous in 2025, experts say. And that includes the open-source versions.

“The rise of multimodal functionality will continue next year,” says Meta’s Choudhury. “Today, most models that mix text, image and speech are an amalgamation of multiple AI models, but this is about to change.” Choudhury predicts that future architectures will be natively multimodal across two or more dimensions, which will lead to some “mind-blowing” new use cases. “Imagine that you can ask a model a question about a picture using just your voice, and it can answer back seamlessly via speech, text or an image. Or how about answering back with all three at the same time?”

Will open-source AI drive more innovation in 2025? Haddad, from the Linux Foundation, believes so. And as AI models become more complex, he believes that organizations will unlock innovation through collaboration with external contributors.

“With open source, developers from all over the world can share their expertise, leading to rapid progress and continuous improvements in gen AI frameworks,” he says. “For instance, in Linux Foundation AI & Data, we have over 100,000 developers contributing to our 68 hosted projects, coming from over 3,000 organizations. Not a single organization out there can match that level of expertise and scale.”

Arnaud Le Hors, a Senior Technical Staff Member of Open Technologies at IBM Research, also envisions collaborations between large enterprises on LLMs becoming more commonplace.

“There is still competition and a race on releasing the most performing models,” Le Hors says. “But we can imagine big corporations collaborating on foundation models in the future.”