Pinterest tells Fortune the company is using open-source AI models to achieve performance similar to that of leading frontier AI models at less than 10% of the cost, particularly for visual and multimodal tasks, along with models it has built in-house for specific purposes. The company solidified the decision in late 2025 and views it as the most pivotal aspect of its AI transformation strategy, citing better personalization for users, the ability to scale faster, and most of all, significant cost savings. 

“We’re a smaller company, so cost matters,” Vicky Gkiza, VP of product management at Pinterest, told Fortune. “We’re trying to be as mindful for the user as well as for the economics.”

Pinterest’s AI pivot

The pivot began with noticing a shift. Gkiza said the company is continuously evaluating different models, and a few months ago, felt the advancement in open-source AI significantly narrowed the performance gap relative to leading closed models.

“Our strategy changed from actually going all in on foundation models to really focusing on open-source and property and compact models for our own use cases,” said Gkiza. “A few months ago, you would not get that type of performance with open-source. So what we see is that the technology is advancing, and we see things that we thought were impossible just a few months ago.”

She said they haven’t completely stopped using frontier models, including using OpenAI for some features and internal use cases like coding. But open-source “is where we are shifting more of our focus and where we see more efficiencies,” she said. In a blog post detailing the strategy, the company referred to it as “a change in technology defaults at Pinterest.” 

In practice, the current process is, identify what the company wants to build, and then identify the best open-source models to support it. For example, Pinterest landed on open-source model Qwen to more deeply analyze the visual content of pins and generate contextual text output. Basically, for images that don’t contain sufficient metadata, Qwen enables the platform to more thoroughly understand the content and provide the right descriptors to power user interactions.

The October launch of Pinterest Assistant, a multimodal chat feature that offers personalized shopping recommendations via images, voice, and text, illustrates how open-source and the company’s custom-built models come together. The core multimodal LLM itself—which oversees the agentic loop and is responsible for query understanding, query planning, and effective tool calling—is open-source. At the same time, the feature also leverages several Pinterest-native technologies that rely on its user and visual foundation models—specifically an underlying set of multimodal retrieval systems, recommendation services, and specialized generative models.

Drivers and benefits

When discussing what drove this strategic shift for Pinterest, Gkiza emphasized that it came down to cost. Relying fully on off-the-shelf models is very expensive, she said, meanwhile the company is citing 90% cost reduction from fine‑tuning open‑source models.

“We didn’t want to reduce our velocity, to reduce the amount of experience that we bring out for our users, but we needed to find a work-around so we’re able to deliver on that,” she said.

In January, Pinterest announced it would lay off 15% of its workforce—an estimated 700 employees—and reduce its office space this year to pivot spending to AI. Companies attributing layoffs to AI have received some pushback, with critics saying they’re using AI as a convenient excuse to cut jobs and costs. 

A Pinterest spokesperson told Fortune the company “made organizational changes to further deliver on our AI-forward strategy, which includes hiring AI-proficient talent. As technology evolves, some roles are transformed, and we want to make sure we approach these changes with empathy and responsibility.”

According to Gkiza, the ability to fine-tune open-source models to meet the company’s specific needs is also allowing Pinterest to move faster, which is further driving savings beyond the upfront benefits of not paying for frontier model usage.

“You get the best of both worlds, and you’re able to scale much faster, much better from an economics perspective,” she said.

As far as benefits, she also cited transparency with open-source models and improved personalization enabled with its proprietary models, which have been trained over many years by user curation. For other companies considering this approach, she emphasizes the importance of focusing on training compact models for specific purposes using the company’s own resources.

“Use the company’s products and data to train these models,” she said. “Because that is how it’s been proven [to get] the best performance in different use cases.”

The ecosystem factor 

Of course, open-source isn’t a golden ticket to cost savings. It comes with its own set of operational nuances and, many argue, a responsibility to contribute back to the ecosystem.

Just as “AI slop” is flooding social media, the same is happening on open-source platforms. Tldraw, a software-development kit project that isn’t open-source but keeps its codebase on GitHub and accepts code contributions, recently closed external contributions for “the good of the project” after becoming overrun with slop contributions late last year, reported LeadDev. In February, GitHub announced new tools to help open-source maintainers combat the increasing onslaught of “low-quality contributions” on the platform, enabled by the efficiencies of AI-assisted coding.

It presents a challenge for companies, which need to carefully review open-source code for vulnerabilities and errors that could wreak havoc in production. It’s not a brand-new issue in open-source, but it is—perhaps ironically—being exacerbated by AI.

And it’s important to remember that an open-source software project’s existence and quality is never guaranteed: These codebases are typically maintained by passionate individuals or small teams, often working without any financial backing. A paper published in January by a team of researchers in Europe titled “Vibe coding kills open-source” describes how AI-powered coding may be lowering the cost of building on existing code, but it’s weakening user engagement through which many maintainers earn returns and “raises a general equilibrium question about the sustainability of open-source software.”

“Sustaining [open-source software] at its current scale under widespread vibe coding requires major changes in how maintainers are paid,” the paper reads.  

Open-source is a two-way street, and well-resourced tech companies in particular sometimes face criticism for taking more than they give to the ecosystem. Regarding how Pinterest is giving back to the open-source community, Gkiza said the company is releasing papers, participating in conferences, and discussing if it should become a more active contributor by sharing some of its models within the open-source community. 

“We have not decided yet,” she said. “But it is something that we’d love to give back to, and we’ll continue evaluating.”