Pinterest Cuts AI Costs by 90% Through Model Customization

Pinterest, serving over 620 million monthly users, faced high costs associated with using advanced AI models for generating recommendations. The solution was a radical redesign of the Qwen3-VL model’s architecture: engineers removed the standard vision layer and integrated custom embeddings trained on the platform’s unique data.

According to CTO Matt Madrigal, the key to success was working with open-source models that can be adapted for specific tasks. Pinterest’s team has long used such solutions—from BERT and CLIP to Qwen3-VL—but now focuses on deep customization. For example, the Navigator 1 shopping assistant model was rebuilt from scratch: instead of a standard vision encoder, custom multimodal embeddings were used, reducing inference latency by 20 times.

Another innovation was the "Taste Graph"—a dynamic system tracking user preferences in real time. Unlike social graphs, it focuses not on connections between users but on their interests, helping convert inspiration into actions like ad clicks or purchases. The model continuously updates based on new data, ensuring a personalized experience.

Madrigal emphasized that for mission-critical functions serving millions of users, the company either develops solutions in-house or maximally adapts open-source models. This approach not only reduces costs but also enhances service quality, which is especially important for a platform centered on visual search and discovery.