Pinterest has replaced a multi-stage clustering system for identifying user projects with a fine-tuned, four-billion-parameter Qwen3 model. Instead of grouping keywords into separate interests, the model reads activity over time and generates a ranked list of broader “journeys,” such as a kitchen renovation that connects cabinets, counters and lighting.
The model can examine up to 150 actions across a 360-day window, along with a compact profile. Search and save activity receives more weight than incidental clicks. Prompt rules discourage merging unrelated interests, suppress one-off behavior and allow an empty result when evidence for a sustained goal is weak.
Pinterest generated synthetic training examples with larger frontier models, then fine-tuned smaller open-weight Qwen3 variants using supervised learning and LoRA. It tested sizes from 0.6 billion to eight billion parameters and chose the 4B model as a balance between quality and serving throughput. Nvidia Dynamo routes requests to vLLM workers on L40S GPUs, while a cache avoids recomputing journeys every day.
In US and Canadian A/B tests, Pinterest reports 1.1% higher email click-through and 1.3% more push opens than the clustering system, plus engagement and retention gains. Those results measure product response, not whether every inferred goal is accurate. The company also uses side-by-side review and a separate model judge to assess relevance.