Integrating Netflix’s Foundation Model into Personalization applications

Divya Gadde, Ko-Jen Hsiao, Dhaval Patel, Moumita Bhattacharya17 Nov 2025

Problem

Netflix's homepage was powered by many specialized models, and each one needed its own time and resources to maintain and improve. The team centralized preference learning in one Foundation Model trained on user interaction histories and content data. The question is how to pass what it learned to the existing application models. Teams differ in latency limits, tech stacks and willingness to pay for a heavy model, so no single integration fits all.

Example from the article: a "Trending now" row cares more about recent interactions with trending titles than about old history, so it needs different adaptation than a title-to-title recommender.

Solution in brief

Netflix runs three integration patterns in production. Pre-computed profile and item embeddings are the cheap default. The model as a subgraph of the application model gives fresh outputs at higher cost. A fine-tuned copy of the model serves a surface directly.

Questions this case answers
  • How to distribute what one large shared model has learned to many downstream models that have different latency limits, tech stacks and compute budgets
  • How to keep pre-computed embeddings usable as features for downstream models when the producing model is retrained and drifts
  • How to reuse one large user-sequence model across many existing recommendation models without retraining each from scratch?
  • How to choose between embeddings, a subgraph and fine-tuning when a foundation model feeds downstream applications?

Tools and methods

  • Profile and item embeddings — hidden state of the last user event and item-tower weights, published daily to an embedding store.
  • Embedding store — in-house feature store for embeddings: versioning, timestamps, offline and online access.
  • Embedding stabilization — maps each day's embeddings into one shared space (see the paper).
  • Subgraph integration — the decoder stack becomes part of the application model graph.
  • Fine-tuning framework — full or partial fine-tuning with an optional new output head.
Continue reading — it's free

Google or a one-time email code. No password.