Two-tower model

Instagram’s Explore system uses two-tower models alongside other methods to find possible recommendations, gathering up to 1,500 photos and videos in its initial selection stage, according to Meta’s Transparency Center1. The Transparency Center also describes a two-tower model in the next stage, where similarity scores help narrow that pool before later selection1. Meta’s engineering team explains that training teaches the model to make those similarity scores predict engagement, such as someone liking a post2. Meta’s engineering team says the design extends an earlier approach based on account and post identifiers, allowing it to use more kinds of information about people, posts and authors and learn several prediction tasks together2.

Meta’s research explores ways to make two-tower models more efficient, without establishing these designs as deployed Instagram features34. A July 2025 Meta research paper describes training on all the items associated with one user request together, so the person-processing part of the model avoids repeated work on shared information, reducing storage and computation and allowing more data and more complex models4. In October 2026, Meta’s engineering team described a two-tower research design whose two sides share a language-processing model, matching conventional two-tower performance using a fraction of the training data while remaining efficient enough to serve recommendations3.