Meta's GEM ads model now runs as one central system across all surfaces

From per-surface systems to one central model

Last year Meta introduced GEM, which it describes as the industry's largest recommendation system (RecSys) foundation model and the company's first LLM-inspired paradigm for ads recommendation. The 2026 generation moves away from isolated optimizations for individual domains such as Facebook, Instagram and Reels. Instead, Meta says, a single central foundation model learns from data and signals across all of its ads platforms.

The company points to the scale of the problem: its apps span dozens of surfaces, with 3.6 billion people using at least one of them each day. Each surface and optimization objective previously had its own independent prediction system. GEM now jointly predicts clicks and conversions across all surfaces and across key advertiser objectives, from off-site conversions to mobile app installs.

Meta says representations learned on data-rich click tasks transfer into data-sparse conversion tasks. A new training recipe with cost-free data mixing and a domain expert framework is meant to make this large, heterogeneous learning tractable.

Beyond the headline gains, Meta reports additional conversion uplift of 5.8% on Facebook, 4.1% on Instagram and 1.6% on off-site conversion for targeted users over the past year.

Kunlun: an LLM-style architecture for recommendations

Meta argues that recommender models have not shown the predictable scaling seen in transformer language models, because of massive heterogeneous features and shifting distributions. To address this, it built an LLM-style architecture called Kunlun, co-designed with the latest GPU hardware to make larger, deeper models more cost-efficient and to support long-context learning over user engagement history.

At its largest pre-training scale, Meta says it observed an LLM-like scaling law with 2.5x model complexity growth from earlier GEM versions. Compared with the previous version, Interformer, Kunlun delivers 2x performance gains with the same computation scaling, according to the company.

Meta lists several design elements:

• Attention: customized attention and feed-forward networks to capture long dependencies in user behavior.
• Block attention: block-aligned sparse attention replaces full self-attention, reducing the complexity tied to sequence length from quadratic to linear relative to window size.
• Generalized Dot-Product Attention (GDPA): Per-Example Feed-Forward Networks are reformulated into attention operations, unifying attention-like operations into a single GPU kernel and delivering up to 2x forward speedup.
• Cross interaction: a mixture-of-experts style design with expert parallelism, so each GPU processes a subset of expert sub-networks in parallel.

Meta also describes a Component Skip Connection (CompSkip) mechanism intended to make ultra-long sequences computationally tractable.

Passing intelligence to user-facing models and training at scale

Serving a model of this size in real time is prohibitively expensive and raises latency, Meta says. Its answer is LoopFM, a post-training technique that transfers GEM's internal representations and sequential historical memory to downstream models. Through computation sharing, Meta says, it doubles the intelligence transfer achieved with the knowledge distillation used in the previous version, at no additional compute cost or latency. The company also introduced elastic pre-computation, called Solaris, to capture gains from the upstream model.

On the training side, Meta notes that data center GPUs and their software stacks are mostly optimized for LLM workloads, while recommendation workloads have a different profile that makes high GPU utilization hard to reach. The company developed custom kernels with low precision and new 5D parallelism strategies, co-designed with the latest GPU hardware and its training cluster.

According to Meta, end-to-end training efficiency doubled to 20-25% model FLOPS utilization (MFU), with 4x total training FLOPs across thousands of GPUs. This, it says, lets GEM train at LLM scale. Meta refers readers to a companion blog, Training GEM at LLM Scale, for the full technical account.

What it means for advertisers

Meta says the result is simultaneous ad performance gains across every surface and funnel stage, from awareness to consideration and conversions, and a deeper understanding that better matches ads to what people are looking for. For advertisers, the company describes the gains as spanning both upper-funnel surfaces and deep-funnel optimization goals, driving stronger returns. All figures come from Meta's own reporting and have not been independently verified.