Pinterest has replaced its multi-step system for grouping user interests with a fine-tuned 4-billion-parameter Qwen3 language model. The company says an initial test in the US and Canada lifted email click-through by 1.1% and push-notification opens by 1.3% compared with its previous system.
Pinterest calls a “journey” a goal that links a person’s interests, intent and circumstances over multiple sessions. For example, searches for kitchen cabinets, saved countertop ideas and later browsing for tiles or lighting may point to one renovation project, rather than separate interests. The model uses those inferred journeys in Pinterest’s existing recommendation system; it does not replace that system.
How does the new system work?
The old pipeline pulled keywords from activity, grouped them using embeddings and hierarchical clustering, then used other models and rules to name and rank the groups. Pinterest says that could split related actions into separate journeys, create vague labels or treat a passing click as lasting intent.
The new model reads a chronological record of up to 150 actions from a 360-day period, plus a compact user profile, and generates ranked journey names. Search and save actions count more than incidental clicks. Its instructions emphasize repeated activity, how recent it is and how specific it appears. They also tell it not to combine unrelated interests, to filter out one-off activity and to return no journey when the evidence is weak.
Pinterest used larger models to create synthetic training examples from activity histories, then fine-tuned smaller, open-weight Qwen3 models. After testing models from 0.6 billion to 8 billion parameters, it chose the 4-billion-parameter version for a balance of inference quality and serving speed. Pinterest runs it with NVIDIA Dynamo and vLLM on L40S GPUs, and avoids rerunning inference for users whose journeys were generated within the previous couple of days.
What did Pinterest measure?
Pinterest evaluates journey quality with side-by-side comparisons, a separate language model that judges whether a journey is relevant and supported by activity, and online experiments. The company reports statistically significant engagement and retention gains in its initial A/B tests, as well as the click and open-rate increases. It did not give the size of the retention gain in the supplied account.
The approach resembles Netflix’s use of long sequences of member activity to improve recommendation systems. Pinterest generates explicit descriptions of journeys, while Netflix has described making reusable representations of preferences; both feed longer-term behavior into existing personalization systems.
Why it matters
For Pinterest users, the change could make project-related emails and push notifications better reflect activity spread across multiple visits; Pinterest reports higher click and open rates in its initial tests. The results do not show whether every user will get more relevant recommendations, or how large the retention improvement was.
