Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI

Meta has announced the 2026 generation of its Generative Ads Model (GEM), a single central Foundation Model designed to unify advertising optimization across Facebook, Instagram, and Reels. Moving away from domain-specific independent optimizations, GEM leverages LLM-scale compute and a new architecture to bend traditional scaling curves, resulting in an 18.9% improvement in conversion rates and a 12% improvement in click-through rates (CTR) across all ad surfaces and objectives.

Kunlun Architecture and Scaling Laws

The core technical advancement is the Kunlun architecture, which co-designs the model with latest GPU hardware to overcome the limitations of traditional recommendation systems. Unlike previous models, Kunlun follows LLM-like scaling laws, achieving a 2.5x increase in model complexity at maximum pre-training scale while delivering 2x performance at the same compute budget compared to the previous Interformer version.

Key architectural innovations include:

• Block Attention: Replaces full self-attention to reduce complexity from quadratic to linear relative to sequence length.
• Generalized Dot-Product Attention (GDPA): Reconfigures pFFN into a single GPU kernel, improving forward speed by up to 2x.
• CompSkip: A compute-skipping mechanism that reduces FLOPS by 43% with minimal quality loss.
• OneSeq: Integrates heterogeneous user sequences chronologically to maximize GPU efficiency.

Intelligence Transfer and Serving Efficiency

To address the high latency and cost of serving large models in real-time, Meta introduced LoopFM, a post-training technique that transfers GEM’s internal representations to downstream models. This approach, combined with knowledge distillation, doubles the effectiveness of intelligence transfer without adding computational overhead.

Additionally, Speculative Precomputation (Solaris) borrows from LLM serving techniques. It asynchronously pre-computes foundation model embeddings for top ad candidates in the background. If exact embeddings are missing, the system uses aggregates from similar users or ad groups, ensuring near-real-time intelligence delivery without dropping requests.

Unified Cross-Domain Training

GEM replaces independent prediction systems across dozens of surfaces and 3.6 billion DAU with a single integrated compute engine. This engine jointly predicts clicks and conversions for all surfaces (Facebook, Instagram, etc.) and objectives (Off-site conversion, Mobile app install).

The training recipe includes:

• Cost-Free Data Mixing: Uses virtual data loading to control mixing ratios without physical data merging, reducing storage costs and overfitting.
• Domain Expert Framework: Utilizes shared experts for unified representations and domain-specific experts for surface-level patterns, leveraging both cross-domain and specialized knowledge.

Training Efficiency and Future Outlook

Meta achieved a 4x increase in total training FLOPs over the past 12 months while doubling End-to-End (E2E) training efficiency to 20-25% MFU. This was accomplished through topology-aware 5D parallelism, custom kernels like Jagged Flash Attention, and MXFP8 mixed ultra-low precision training.

Future plans include expanding GEM to learn from all platform data (organic and ads, text, image, video) and building LLM-scale long contexts to enhance targeting efficiency and ad relevance.