Pinterest cut AI costs 90% by gutting a frontier model's vision layer

Pinterest CTO Matt Madrigal said the company customized Qwen3-VL for its Navigator 1 shopping assistant by removing the model’s vision encoder and using Pinterest’s proprietary multimodal embeddings instead. The change reportedly cut AI costs by 90%, improved accuracy by 30%, and avoided runtime image encoding that would have made inference latency far worse.

The move shows how companies with large, unique datasets can make open-source models more efficient than generic frontier models for specialized products. It also signals a broader shift toward enterprise AI architectures built around proprietary data pipelines and precomputed embeddings rather than simply scaling model calls.