Source link : https://tech365.info/pinterest-reduce-ai-prices-90-by-gutting-a-frontier-mannequins-imaginative-and-prescient-layer/
At 620 million month-to-month customers, calling a frontier mannequin for each picture advice isn’t a method — it’s a invoice. Pinterest CTO Matt Madrigal solved it by gutting Qwen3-VL’s imaginative and prescient layer and rebuilding it with proprietary embeddings, chopping prices 90% and boosting accuracy 30%.
Madrigal’s group has been closely investing in customizing open-source fashions “foundationally in-house.”
“If you’ve got really unique data that you can then fine-tune an open source model with, data quality will, frankly, outweigh or overcome model size,” Madrigal defined in a current VB Past the Pilot podcast.
How Pinterest custom-made Qwen for visible discovery
Pinterest, which has round 620 million month-to-month energetic customers, has lengthy utilized open supply fashions for visible search and discovery, going again to Google’s BERT and OpenAI’s CLIP. The corporate fine-tuned its personal Pin CLIP on the latter, incorporating proprietary visible embeddings and picture metadata.
Pinterest’s conversational buying assistant, Navigator 1, was constructed on Qwen3-VL and customised in “pretty significant” methods. Madrigal’s group basically “ripped out” Qwen’s imaginative and prescient encoder layer and fine-tuned the mannequin on proprietary multimodal embeddings. This has allowed them to seize metadata round pins and pictures that may then be precomputed offline and recurrently retrained on new info to ship personalised…
—-
Author : tech365
Publish date : 2026-05-29 18:51:00
Copyright for syndicated content belongs to the linked Source.
—-
1 – 2 – 3 – 4 – 5 – 6 – 7 – 8