Pinterest cut AI costs 90% by gutting a frontier model's vision layer
At 620 million monthly users, calling a frontier model for every image recommendation isn't a strategy — it's a bill. Pinterest CTO Matt Madrigal solved it by gutting Qwen3-VL's vision layer and rebuilding it with proprietary embeddings, cutting costs 90% and boosting accuracy 30%. Madrigal’s team has been heavily investing in customizing open-source models “foundationally in-house.” “If you've got really unique data that you can then fine-tune an open source model with, data
Pinterest has significantly reduced its artificial intelligence costs by customizing open-source models. The company's CTO, Matt Madrigal, detailed how his team modified the Qwen3-VL model for their conversational shopping assistant, Navigator 1. By replacing the model's original vision layer with proprietary embeddings and fine-tuning it on unique data, Pinterest achieved a 90% cost reduction and a 30% improvement in accuracy.
This approach leverages Pinterest's extensive proprietary data to enhance open-source foundation models. The company has a history of customizing models like BERT and CLIP to improve visual search and discovery for its 620 million monthly users. This strategy allows for more efficient and personalized user experiences.
This innovation demonstrates a practical and cost-effective method for large platforms to leverage advanced AI by tailoring open-source solutions to their specific needs, potentially influencing how other companies approach AI development and cost management.
📌 Kaynak
Bu özet VentureBeat kaynağından otomatik derlenmiştir. Tamamı için orijinal habere gidin.
Orijinal haberi oku →