Meta’s SilverTorch Unifies Recommendation Retrieval into a Single Neural Network

Meta has introduced SilverTorch, a new recommendation system that replaces traditional microservice-based retrieval with a unified neural network architecture called Index as Model. The system consolidates all retrieval components—approximate nearest neighbor search, eligibility filtering, neural reranking, and multi-task scoring—into a single PyTorch model. In an 80-million-item evaluation, SilverTorch achieved 23.7× higher throughput and 20.9× better compute cost efficiency compared to CPU-based baselines, while also improving recommendation quality. The research paper has been accepted as a full paper at SIGIR 2026. SilverTorch is already deployed across Meta’s family of apps, including Facebook, Instagram, and Threads, and is designed to integrate large language models as additional modules. The system uses streaming updates for index freshness, enabling same-day posts to constitute a significant portion of recommendations. Key technical innovations include a fused Int8 ANN search kernel and a Bloom index filter, both reimplemented in pure PyTorch to leverage GPU parallelism.