Meta Expands Generative Ads Model GEM to Scale Ads AI
Meta has shared details on the latest iteration of its Generative Ads Model (GEM), shifting from isolated domain optimizations to a centralized foundation model designed to scale compute and intelligence transfer.
Radwan AbdulhadiEditor and technology writer
Transitioning to a Central Foundation Model
Last year, Meta introduced GEM as the industry’s largest recommendation system foundation model, establishing a large language model-inspired paradigm for ads recommendation. Building on that foundation, the 2026 generation of GEM shifts away from isolated per-domain optimizations across platforms like Facebook, Instagram, and Reels, unifying them into a single central foundation model.
According to technical disclosures from Meta Engineering, the updated system learns from signals across the entire ads ecosystem, utilizing post-training techniques to provide a larger latency budget without requiring real-time evaluation of all ad candidates on an efficient training stack.
Performance and Business Metrics
Meta reports that GEM has delivered tangible performance gains across upper-funnel surfaces and deep-funnel optimization goals. Over the past year, the model achieved an 18.9% conversion improvement and a 12% ad click-through rate (CTR) improvement across its ad surfaces and objectives.
Additionally, the system contributed specific conversion uplifts of 5.8% on Facebook, 4.1% on Instagram, and 1.6% for off-site conversions among targeted users, driving stronger overall returns for advertisers.
Bending the Scaling Curve With Kunlun
Traditional recommender system models typically fail to exhibit the predictable scaling properties seen in transformers due to massive heterogeneous features and shifting data distributions. To address this challenge, Meta built an LLM-style architecture named Kunlun, co-designed with modern GPU hardware.
This hardware-software co-design makes larger, deeper models more cost-efficient while facilitating long-context learning over user engagement history. Meta observed an LLM-like scaling law at pre-training scale, achieving a 2.5x model complexity growth compared to previous iterations.
Efficient Intelligence Transfer and Post-Training
Serving a massive recommendation foundation model in real time presents significant latency challenges that can compromise user experience and advertiser return on investment. To counter this, Meta implemented structured representations through LoopFM, a post-training technique that transfers internal representations and sequential historical memory to downstream models.
Through computation sharing, this approach doubles intelligence transfer relative to previous knowledge distillation methods without imposing additional compute cost or latency. Meta also introduced elastic pre-computation via Solaris to capture upstream model gains.
Training Scale and Infrastructure Efficiency
Training a recommendation foundation model across thousands of GPUs requires specialized orchestration of compute, parallelism, communication, and memory. Because standard data center software stacks are primarily optimized for LLM workloads, recommendation profiles demand distinct approaches.
Meta developed custom low-precision kernels alongside new 5D parallelism strategies co-designed with cluster hardware. These infrastructure optimizations doubled end-to-end training efficiency to 20 to 25 percent model FLOPS utilization while scaling total training FLOPs by 4x.
Sources
- Meta EngineeringMeta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI
Continue chronologically
Related entity coverage
Related stories
Meta Expands GEM RecSys Model to Scale Ads AI
Meta details the evolution of its Generative Ads Model (GEM), scaling compute and intelligence transfer across its global ads ecosystem.
Meta Introduces F3 Lightning to Scale Ads ML Features
Meta details F3 Lightning, a new infrastructure layer that accelerates ads machine learning features to production without increasing costs.
Meta Transforms Ads Recommendation with GenRecs Tokenization
Meta introduces GenRecs Tokenization, a foundational data representation leveraging stable semantic vocabularies to power generative ad recommendations.
Meta Engineers Detail Thin Orchestrator AI Agent Architecture
Meta details a thin orchestrator AI agent architecture that improves efficiency and reliability by giving central routers less domain work.
