Meta's GEM 2026: One Central Foundation Model for Ads Recommendation

Meta Scales GEM Into a Central Ads Model

Meta engineering has detailed the 2026 generation of GEM, its Generative Ads Model, which shifts from isolated per-domain optimizations to a single central foundation model for ads recommendation. GEM learns from data and signals across all of Meta's ads platforms and delivers full-funnel optimization spanning awareness, consideration and conversions.

The company reports that GEM has delivered an 18.9% conversion improvement and a 12% ad click-through rate improvement across Meta's ad surfaces and objectives, along with 5.8%, 4.1% and 1.6% conversion uplift on Facebook, Instagram and off-site conversion for targeted users over the past year.

Kunlun Architecture Bends the Scaling Curve

To make larger models cost-efficient, Meta built Kunlun, an LLM-style architecture co-designed with the latest GPU hardware, enabling an LLM-like scaling law with 2.5x model complexity growth from earlier GEM versions. Its standardized layers combine customized attention, a summarization component and cross interaction using a mixture-of-experts style design, while block-aligned sparse attention cuts sequence complexity from quadratic to linear relative to window size, and a reformulated per-example feed-forward network delivers up to 2x forward speedup.

Sequence efficiency comes from CompSkip, a computation skip mechanism that cuts FLOPS by 43% with negligible quality loss, and OneSeq, which unifies heterogeneous user sequences into a single processing layer. GEM scaled its long engagement context length by 2x in production and demonstrated infrastructure capability to scale long context to 25x, finding that prediction quality follows a power law with sequence length and that larger models benefit disproportionately from longer sequences.

Intelligence Transfer and Cross-Surface Learning

Because serving the foundation model in real time is too expensive, GEM transfers intelligence downstream through LoopFM, a structured representation technique that pairs deep historical memory with knowledge distillation to double transfer effectiveness at no extra compute cost. Solaris adds speculative precomputation, evaluating likely user-ad pairs in asynchronous background jobs just before vertical model inference needs them, with hierarchical fallbacks to aggregate or similar-user embeddings when exact matches are unavailable.

At the systems level, GEM unifies prediction across dozens of surfaces used by 3.6 billion people daily, jointly predicting clicks and conversions for objectives from off-site conversions to mobile app installs. A cost-free data mixing strategy and a domain expert framework with shared and surface-specific experts make this heterogeneous learning tractable.

Training at LLM Scale

GEM now trains on several thousand latest-generation GPUs, with custom kernels such as Jagged Flash Attention and mixed ultra-low precision training including MXFP8 for attention and MLP components, alongside topology-aware 5D parallelism co-designed with Meta's network hierarchy. Meta reports that end-to-end training efficiency doubled to 20-25% model FLOPS utilization over 12 months while total training FLOPs scaled by 4x.

Looking ahead, Meta plans to extend GEM to learn from all platform data including text, image and video engagements, push toward LLM-scale long context, scale on larger compute clusters and introduce AI-driven optimizations such as architecture co-design, evolving post-training with reasoning capability to transfer more intelligence to downstream models.

Sources: Engineering At Meta · Meta Engineering