Netflix Tests GenRec LLM to Power Recommendations

Netflix has developed GenRec, a recommendation system powered by a roughly ten-billion-parameter open-weight language model. Instead of relying on thousands of hand-crafted numerical features to represent user interactions, GenRec translates user history—such as watch durations, thumbs up or down, and drop-offs—into plain text. The system runs on vLLM, utilizing a specialized mode that scores all catalog candidates in a single pass without generating text, which keeps compute costs manageable.

The training process occurs in two distinct phases. First, Netflix fine-tunes the base open-weight model on its catalog and user data. Second, it applies specialized training to turn the model into a recommendation ranker. This second phase adds 35 to 50 percent to the base model's performance, a gap that widens to 80 percent when the base model is two weeks old. To prevent the model from hallucinating non-existent titles, Netflix integrates a separate component that restricts recommendations to actual catalog entries.

In offline evaluations, GenRec delivered a 1.6 percent improvement in ranking quality over Netflix's highly tuned production system, while requiring roughly 40 times fewer labeled examples during its second training stage. Netflix also conducted a four-week online A/B test across approximately ten percent of its traffic. The live experiment yielded a 0.115 percent increase in a short-term home screen metric and a 0.006 percent rise in a long-term core metric, both of which are statistically significant.

For machine learning practitioners, GenRec represents a fundamental shift from manual feature engineering to context engineering. Instead of building custom architectures for every recommendation task, a single generative model can handle multiple use cases, making it much easier to onboard new content types like podcasts, games, or live streams. Netflix views this as part of an industry-wide transition toward general-purpose language models, aligning with other recent research such as PLUM, GLIDE, and OneRec-Think.