Meta reports gains from new AI-powered ad recommendation model GEM

Meta has detailed the progress of its Generative Ads Recommendation Model (GEM), its artificial intelligence (AI)-driven recommendation system for ads.

The foundation model aims to enhance how the company personalises ad delivery by learning from large-scale data and interactions. According to the company, since its rollout earlier this year, GEM has contributed to a 5% increase in ad conversions on Instagram and a 3% increase on Facebook Feed during the second quarter of 2025. The company also said architectural improvements in the third quarter doubled the performance benefit achieved from data and compute scaling, allowing for more efficient training and returns.

It said GEM, which draws on large language model (LLM) principles and runs on thousands of GPUs, is designed to improve ad relevance and performance across Facebook and Instagram.

GEM represents a new approach to ad recommendation systems, introducing architectural and training innovations that enable more efficient scaling. It incorporates techniques such as multi-dimensional parallelism, custom GPU kernels, and memory optimizations to train at a scale comparable to large language models. The model’s design allows it to recognise complex patterns across diverse ad and user data, delivering more personalized and relevant ad recommendations.

To address the challenge of managing large, dynamic datasets across Meta’s platforms, GEM processes vast amounts of ad-related data, including advertiser goals, creative formats, and user behaviour, while maintaining efficiency across thousands of GPUs. The company said the model’s architecture is four times more efficient than previous versions and achieves double the effectiveness in knowledge transfer compared to standard methods.

The system learns from user interactions across the company’s apps, including both organic and paid content. It models these behaviours using customised attention mechanisms to understand long-term engagement patterns and improve ad targeting accuracy. The model also applies domain-specific optimisation, learning from cross-platform activity, such as Instagram video engagement, to refine predictions on other surfaces like Facebook Feed.

To ensure that GEM’s improvements benefit multiple models across the company’s advertising stack, it uses a layered knowledge transfer framework. This includes direct and hierarchical transfer methods that help propagate the foundation model’s learnings to other ad systems, improving efficiency and consistency. Techniques like representation learning and parameter sharing further streamline this process.

Training GEM at such a scale required a redesign of the company’s infrastructure to optimise GPU efficiency. The company implemented distributed training strategies, hybrid sharding methods, and custom GPU kernels to manage variable-length user data and improve processing speed. The company said these efforts led to a 23-fold increase in training performance and a 1.43-times improvement in hardware utilisation.

The company also reduced job startup times by fivefold and improved compilation speed through caching strategies in PyTorch 2.0. These improvements are part of a broader focus on enhancing GPU utilisation across the full model lifecycle, from early experimentation to large-scale deployment and post-training tasks.

Looking ahead, the company plans to expand GEM’s capabilities to learn from multimodal data across text, images, audio, and video. The company said future iterations will cover all major surfaces across Facebook and Instagram, supporting a unified model that can rank both ads and organic content.

Meta said the goal is to develop a deeper understanding of user preferences and intent to make ad experiences more personal, while also providing advertisers with stronger engagement outcomes. The company plans to continue scaling GEM on larger AI clusters and refining its architecture for improved efficiency and predictive precision.