Pinterest has replaced its keyword-clustering system for understanding user behavior with a fine-tuned 4-billion-parameter language model, delivering a 1.1% improvement in email click-through rates and 1.3% higher push notification opens, according to a report published by InfoQ this week. The shift moves the company's user-journey inference from a multi-stage pipeline that grouped keywords into interests to a model that examines activity histories to identify longer-term goals. Those journeys then feed into Pinterest's existing recommendation infrastructure.
The company defines a "journey" as a goal connecting interests, intent, and context across multiple sessions. A person renovating a kitchen, for instance, might look up cabinet designs, bookmark countertop concepts, and later browse backsplash tiles or lighting fixtures—interactions that collectively point to a single coherent project rather than disconnected interests, the report notes. The previous system pulled keywords from user activity, clustered them using embeddings and hierarchical grouping, labeled the resulting clusters, and ranked them with additional models and diversity rules. While effective compared with earlier interest-based notifications, the approach could split related activities into separate journeys, generate vague labels, or confuse incidental engagement with sustained intent.
The new system treats journey inference as a generation task. A fine-tuned language model reads a chronological record of up to 150 user actions spanning a 360-day window, plus a condensed user profile, and produces a ranked list of journey names. Search and save actions receive more weight than casual clicks, while the prompt directs the model to weigh repeated engagement, recency, and specificity. The consolidation shifted much of the earlier system's ranking and filtering logic into the prompt, according to the report. Instructions discourage merging unrelated interests, suppress one-time activity and noisy auto-generated tags, and permit an empty result when evidence of a meaningful journey is insufficient. Pinterest treats changes to these instructions as production launches rather than simple prompt edits. To build the system without extensive human-labeled data, Pinterest used frontier models to generate synthetic training examples from user activity histories, then fine-tuned smaller open-weight Qwen3 models through supervised fine-tuning and low-rank adaptation. The team tested models from 0.6 billion to 8 billion parameters, selecting the 4-billion model for its balance of inference quality and serving throughput.
Running the model at Pinterest's scale posed a separate engineering challenge. The company deployed NVIDIA Dynamo in front of vLLM workers on L40S GPUs, using continuous batching, PagedAttention, and prefix-cache-aware routing to boost inference efficiency. One of the largest throughput gains, however, came from skipping unnecessary inference altogether: the system tracks when each user's journeys were last generated and bypasses users processed within the previous couple of days. Pinterest evaluates journey quality through side-by-side comparisons, a separate language-model judge assessing relevance and support from user activity, and online experiments. Initial A/B testing in the United States and Canada produced the reported gains in email click-through and push opens, alongside statistically significant improvements in engagement and retention. The approach shares similarities with Netflix's work on foundation models for personalized recommendations, which learns from long sequences of member interactions to create representations reused by downstream recommendation models. While Netflix focuses on reusable preference representations and Pinterest generates explicit journey descriptions, both use longer-term behavioral histories to enrich existing personalization infrastructure rather than replace it entirely. As organizations navigate the trade-off between complex multi-stage pipelines and end-to-end model architectures, the choice increasingly hinges on whether infrastructure can absorb latency and whether teams can maintain prompt governance at production rigor.