Artificial Intelligence

Meta Expands GEM RecSys Model to Scale Ads AI

Meta has detailed the latest evolution of its Generative Ads Model (GEM), scaling its central recommendation foundation model and compute engine across its global advertising ecosystem.

Radwan AbdulhadiEditor and technology writer

Published: October 13, 2026 · Updated: October 13, 2026 · 3 min read

Meta Expands GEM RecSys Model to Scale Ads AI — Image: Meta Engineering

Evolution of Meta's Central Ads Foundation Model

Meta has provided an engineering update on its Generative Ads Model, known as GEM, which serves as the industry's largest recommendation system foundation model. Initially introduced to establish an LLM-inspired paradigm for ads recommendation, the platform has scaled compute and intelligence transfer to cover Meta’s entire ads ecosystem. You can read the original announcement details regarding the foundation in the initial release of GEM.

The 2026 generation of GEM transitions away from isolated per-domain optimizations across individual applications like Facebook, Instagram, and Reels. Instead, it relies on a single central foundation model that learns from signals across all of Meta’s advertising platforms. According to Meta Engineering, this approach uses LLM-scale computation to deliver post-training techniques that allow a larger latency budget without requiring real-time evaluation of every single ad candidate.

Performance Gains and Business Results

The scaled deployment of the recommendation system foundation model has driven significant performance milestones across multiple surfaces and optimization objectives. Over the past year, the architecture contributed notable conversion uplifts for targeted users on Facebook, Instagram, and off-site conversion metrics.

In total, Meta reported that GEM has achieved an 18.9% conversion improvement alongside a 12% ad click-through rate (CTR) improvement across its diverse ad surfaces and objectives. These optimization gains span both upper-funnel awareness campaigns and deep-funnel conversion goals.

Bending the Scaling Curve with Kunlun

Traditional recommender systems historically struggled to exhibit the predictable scaling properties seen in large language models due to massive heterogeneous features and shifting distributions. To solve this challenge, Meta built an LLM-style architecture named Kunlun, which was co-designed with modern GPU hardware to make deeper models more cost-efficient.

The Kunlun architecture enables long-context learning over user engagement history. At its largest pre-training scale, Meta observed a 2.5x growth in model complexity compared to earlier versions of GEM. This design standardizes the architecture framework, interleaving lifelong sequence modeling with scalable recommendation-specific cross-interaction.

Efficient Intelligence Transfer and Post-Training Techniques

Serving a massive recommendation model in real time presents severe latency and cost challenges. To address this, Meta incorporated a structured representation known as LoopFM, which functions as a state-of-the-art post-training technique designed to transfer internal representations and sequential historical memory to downstream models.

Through computation sharing, LoopFM doubles intelligence transfer alongside knowledge distillation methods from previous iterations, achieving this without adding extra compute cost or latency. Additionally, Meta introduced elastic pre-computation, branded as Solaris, to capture upstream model gains.

Compute Infrastructure and Training Efficiency

Training a foundational recommendation model across thousands of GPUs requires specialized orchestration of parallelism, communication, and memory. Because standard data center software stacks are primarily optimized for LLM workloads, recommendation workloads present distinct profiling hurdles.

Meta developed custom low-precision kernels and 5D parallelism strategies co-designed with its hardware cluster. These enhancements successfully scaled total training FLOPs by 4x while doubling end-to-end training efficiency to achieve between 20% and 25% model FLOPS utilization.