Pinterest switches to fine-tuned Qwen3 for journeys

Pinterest replaced its user-journey inference system with a fine-tuned 4-billion-parameter Qwen3 language model, resulting in a 1.1% improvement in email click-through rates and 1.3% higher push opens. The system processes up to 150 user actions across a 360-day window to generate ranked journey names, reducing reliance on multi-stage clustering pipelines. To operate at scale, Pinterest deployed NVIDIA Dynamo in front of vLLM workers on L40S GPUs and implemented caching to skip recently processed users. Open article →