Inside GEM, Meta’s ads foundation model that boosts conversions

New Delhi: Meta has shared the details foundation model for advertising, called the Generative Ads Model (GEM), which the company describes as the “central brain” behind its ad recommendation system. 

Built with large-language-model-scale techniques and trained on thousands of GPUs, GEM is already contributing to higher ad conversions across Instagram and Facebook, Meta said in an engineering blog post on November 10. 

According to Meta, GEM is a single, LLM-inspired model that boosts the performance of other ads models by improving relevance and prediction quality. Since launching earlier this year, GEM has helped deliver about a 5% increase in ad conversions on Instagram and a 3% increase on Facebook Feed in Q2. 

Meta added that subsequent architecture changes in Q3 doubled the performance benefit per unit of data and compute, supporting further scale-up at attractive returns. 

How GEM works

Meta positions GEM as a foundation model for recommendation, rather than a standalone ranker. It learns from diverse ads and organic signals, then transfers that knowledge to surface-specific models across Facebook and Instagram. 

The company highlights three pillars: a scalable architecture designed to keep improving with more parameters, a suite of post-training techniques to transfer knowledge efficiently to downstream models, and a training stack that can utilise thousands of GPUs with higher throughput.

On efficiency, Meta claims GEM delivers roughly four times the ad-performance gain per unit of data and compute compared with its previous generation of models. 

Knowledge transfer to downstream systems is said to be twice as effective as standard distillation methods. System-level changes have raised effective training FLOPS and improved utilisation so the model can iterate faster at scale. 

Under the hood, GEM processes two broad input types: long user-behaviour sequences and non-sequence attributes such as user profile, ad format and creative features. Meta says it uses customised attention for each, plus cross-feature learning so the model preserves full sequence information while capturing interactions across features. 

The post references components such as an enhanced “Wukong” block for non-sequence interactions and an “InterFormer” design to alternate sequence learning with cross-feature layers. 

What it means for advertisers

Meta frames GEM as a way to jointly optimise user and advertiser objectives across the funnel—from awareness to conversion. In near-term terms, that is intended to translate into more relevant impressions, better conversion rates and improved return on ad spend. Longer term, Meta says GEM will learn across modalities—text, image, audio and video—so it can better capture the nuances behind clicks and long-term value, and ultimately help unify ranking for both organic content and ads. 

The company also ties GEM to faster iteration: by centralising learning and then propagating it to the rest of the ads stack, new signals or improvements can, in theory, lift many downstream models at once rather than being rebuilt surface by surface. Meta notes the model was built to support continuous online training and post-training knowledge generation, with scheduling designed to balance compute loads efficiently.

Context: Meta’s broader ads AI push

GEM follows earlier infrastructure upgrades such as Andromeda, a personalised ads retrieval engine that Meta said improved recall and ad quality at scale. Together, these systems are part of a multi-year effort to automate more of campaign setup, retrieval and ranking, while layering in generative tools on the creative side. 

The road ahead

Meta says it will keep scaling GEM on larger clusters and broaden coverage to all major surfaces across Facebook and Instagram. The goal is a stronger multi-modal foundation that can power “intent-centric” user journeys and enable more agentic automation for advertisers. In short, the company is betting that a single, ever-learning model can raise baseline performance for the entire ads system and accelerate new features into production.