“Build or buy” has been a perennial question for companies looking to integrate AI models, from the early machine learning days several years ago, and still today in the midst of the generative AI boom. While off-the-shelf options are stronger than ever, Pinterest is one company with a firm stake in the former.
Pinterest tells us luck The company uses open-source AI models, along with purpose-built in-house models, to achieve performance similar to leading frontier AI models at less than 10% of the cost, especially on visual and multimodal tasks. The company solidified this decision in late 2025 and considers it the most important aspect of its AI transformation strategy, citing improved user personalization, the ability to scale faster, and, above all, significant cost savings.
“We’re a small company, so cost is important to us,” says Vicky Gkiza, Pinterest’s vice president of product management. luck. “We try to look out for the user as well as the economics.”
Pinterest’s AI Pivot
The pivot started with noticing change. Gkiza said the company continually evaluates different models and several months ago felt that advances in open source AI had significantly narrowed the performance gap with leading closed models.
“Our strategy has changed from really going all in on the underlying model to focusing on compact models for open source, properties, and unique use cases,” Gkiza said. “A few months ago, you couldn’t get that kind of performance with open source. So the technology is advancing and we’re doing things that seemed impossible just a few months ago.”
He said the company has not completely stopped using frontier models, such as using OpenAI for some functions and internal use cases such as coding. But open source is “where we’re shifting our focus and where we think we can be more efficient,” she said. In a blog post detailing the strategy, the company called it a “default change in technology at Pinterest.”
In practice, the current process is to identify what a company wants to build and then identify the best open source model to support it. For example, Pinterest adopted the open source model Queen to better analyze the visual content of Pins and generate contextual text output. Essentially, for images that don’t have enough metadata, Queen allows the platform to more thoroughly understand the content and provide appropriate descriptors to enhance user interaction.
The October release of Pinterest Assistant, a multimodal chat feature that provides personalized shopping recommendations through images, audio, and text, shows how open source models and the company’s custom-built models work together. The core multimodal LLM itself, which oversees the agent loop and is responsible for understanding queries, planning queries, and invoking effective tools, is open source. At the same time, this feature also leverages several Pinterest native technologies that rely on user and visual foundation models, particularly a multimodal search system, recommendation services, and an underlying set of specialized generative models.
Drivers and benefits
When discussing what drove this strategic shift for Pinterest, Gkiza emphasized that it comes down to cost. Relying completely on off-the-shelf models would be prohibitively expensive, she said, but the company can reduce costs by 90% by tweaking open-source models.
“We didn’t want to slow things down to reduce the amount of experience we provide our users, but we had to find a workaround to be able to do that,” she said.
In January, Pinterest announced it would lay off 15% of its workforce (an estimated 700 people) and reduce office space this year to redirect spending toward AI. Companies that blame AI for layoffs have also faced some backlash, with critics accusing them of using AI as a convenient excuse to cut staff and cut costs.
A Pinterest spokesperson said: luck The company said, “We have made organizational changes to further execute on our AI acceleration strategy, including hiring talent skilled in AI. As technology evolves, some roles will change. We want to ensure we approach these changes with empathy and responsibility.”
According to Gkiza, the ability to fine-tune the open source model to the company’s specific needs allows Pinterest to move more quickly, driving savings beyond the upfront benefits of not paying for the Frontier model.
“You get the best of both worlds and can scale faster and better from an economic standpoint,” she said.
As a benefit, he also mentioned the transparency that comes with open-source models and the increased personalization that comes with proprietary models that have been trained over many years through 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 your company’s products and data to train these models,” she said. “Because it’s been proven [to get] Achieve the best performance for a variety of use cases. ”
ecological factors
Of course, open source is not the golden ticket to cost savings. It comes with its own operational nuances and many argue that it has a responsibility to contribute to the ecosystem.
Just as “AI slop” is all over social media, the same thing is happening on open source platforms. Tldraw, a software development kit project that is not open source but maintains its codebase on GitHub and accepts code contributions, recently ended external contributions “for the benefit of the project” after being overwhelmed with contributions late last year, LeadDev reported. In February, GitHub announced new tools to help open source maintainers combat the growing number of “poor-quality contributions” attacks on the platform through AI-assisted coding efficiencies.
This poses a challenge for companies that need to carefully review open source code for vulnerabilities and errors that can wreak havoc on their production environments. This is not an entirely new problem in open source, but perhaps ironically, it has been exacerbated by AI.
And it’s important to remember that the existence and quality of open source software projects is never guaranteed. These codebases are typically maintained by passionate individuals or small teams, often working without financial support. A paper titled “Vibe Coding Kills Open Source” published by a team of European researchers in January explains how AI-enabled coding may reduce the cost of building existing code, but it also weakens the user engagement that many maintainers earn from it, “raising general equilibrium questions about the sustainability of open source software.”
“last [open-source software] “At the current scale of Vibecoding’s prevalence, significant changes are needed in how maintainers are paid,” the paper reads.
Open source is a two-way street, and technology companies, especially those with sufficient resources, can sometimes face criticism for taking more from the ecosystem than they give. As for how Pinterest is giving back to the open source community, Gkiza said the company is discussing whether to become a more active contributor by publishing papers, attending conferences, and sharing some of its models within the open source community.
“I haven’t decided yet,” she said. “But it’s something we would love to give back and continue to evaluate.”
