AI News

Trending story

Meta introduces its ad delivery system Intelligent Delivery

1 reports1 story updatesUpdated Nov 12 01:00

Understand this story

AI overview

On November 12, 2026, the Meta Engineering Blog published a post introducing its ad delivery system, Intelligent Delivery. The system is driven by machine learning and progressively incorporates agentic AI to dynamically allocate compute in ad ranking. According to the article, from 2017 through the first half of 2026, the system contributed an additional roughly 10% improvement in ad ranking performance on top of advances in models and systems. Its Resource Core Model contains 6B sparse parameters and 20M dense parameters, runs before candidate retrieval, and is shared across the entire delivery funnel; on offline benchmarks, its ability to identify high-value ad requests improved by about 20% compared with the previous production model. The system directs compute to the locations with the highest marginal utility, by request, model, hardware cluster, and ranking stage.

AI-generated from public reports

Latest development · Report

Meta Intelligent Delivery: Using ML to Dynamically Allocate Compute for Ad Ranking, Delivering About a 10% Ranking Performance Improvement

Developments

1 confirmed developments

  1. Meta introduces its ad delivery system Intelligent Delivery

    Source:Meta Engineering Blog

Reports and discussion timeline

Browse reports and discussion by time. Their evidence levels remain distinct.

Nov 12

  1. Meta Engineering BlogFirst-party

    Meta Intelligent Delivery: Using ML to Dynamically Allocate Compute for Ad Ranking, Delivering About a 10% Ranking Performance Improvement

    Meta built Intelligent Delivery, a compute allocation system driven by ML that progressively introduces agentic AI, contributing an additional ~10% improvement in ad ranking performance from 2017 through the first half of 2026 on top of model and system advances. Its Resource Core Model contains 6B sparse parameters and 20M dense parameters, runs before candidate retrieval, and is shared across the entire delivery funnel; on offline benchmarks, its ability to identify high-value ad requests improved by about 20% over the previous production model. The system directs compute to the locations with the highest marginal utility across requests, models, hardware clusters, and ranking stages.

Related stories

  1. Related storiesMeta updates its commercial terms, removing language that could have allowed scraping of news website content
  2. Related storiesAnthropic releases Claude Opus 5.5
  3. Related storiesBaidu launches large-model service trading platform "Ciyuansuo"
  4. Related storiesDatabricks describes how Snorkel AI uses Lakebase branches to support CI/CD