Pinterest has achieved significant cost savings in its artificial intelligence operations by extensively customizing an open-source model rather than relying on expensive frontier models for its 620 million monthly users. The company's Chief Technology Officer Matt Madrigal reported that his team reduced AI costs by 90% while simultaneously improving accuracy by 30%.

The social media platform accomplished this by modifying Qwen3-VL, an open-source vision-language model, removing its original vision encoder layer and replacing it with Pinterest's proprietary visual embeddings and image metadata. This approach allows Pinterest to precompute visual data offline and retrain models regularly on new information, avoiding the need to process each image individually at runtime.

Madrigal explained that without these custom embeddings, developers would need to encode each returned image separately during runtime, resulting in latency that is "20 times worse" from an inference perspective. The customized model now powers Pinterest's conversational shopping assistant, Navigator 1, and supports the platform's visual search and discovery features.

The company has also developed what it calls a "taste graph" - a dynamic representation of individual user preferences that goes beyond simple click behavior. This system captures billions of users' evolving tastes and guides them from inspiration to purchase intent through what Pinterest terms "lateral exploration." The architecture combines graph structures with representational learning, continuously updating user embeddings based on activity and new content.

Pinterest's approach reflects a broader trend among enterprises seeking to balance AI capabilities with cost efficiency. Rather than relying entirely on expensive third-party models, the company has invested heavily in customizing open-source alternatives with Apache licenses that allow modification of model weights for specific use cases.