Meta's GEM now runs every ad auction with one model trained across Facebook, Instagram and Reels

The 2026 generation of Meta's Generative Ads Model (GEM) shifts from isolated per-domain optimizations (such as Facebook, Instagram, Reels) to a single central foundation model that learns from data and signals across all of Meta's ads platforms, according to an engineering post published by the company on 13 October.

Meta reports that GEM has delivered 18.9% conversion improvement and 12% ad click-through rate (CTR) improvement across its ad surfaces and objectives, and additionally contributed 5.8%, 4.1%, and 1.6% conversion uplift on Facebook, Instagram, and off-site conversion for targeted users over the past year.

The figures come only from Meta, with no baseline or method attached. They describe what Meta says changed, not something an advertiser can confirm in their own account by isolating GEM's contribution from other system updates that shipped during the same period.

What GEM actually does (and what it does not do)

The "generative" in GEM refers to generating predictions about which ads will perform for which users. It does not generate ad creative. The model sits upstream of the delivery systems that actually serve ads in real time, acting as what Meta calls a "central brain" that transfers learned intelligence to ranking and retrieval engines like Lattice and Andromeda.

GEM is Meta's most advanced ads foundation model, built on an LLM-inspired paradigm and trained across thousands of GPUs, described as the largest foundation model for recommendation systems (RecSys) in the industry, trained at the scale of large language models.

The practical distinction matters. GEM is too computationally expensive to run in real time during an ad auction, which needs to complete in under 100 milliseconds. GEM transfers its learning into Lattice, Andromeda, and the downstream vertical models through a process called knowledge distillation.

Meta's new Kunlun architecture, co-designed with the latest GPU hardware, delivers 2x performance gains over the prior Interformer architecture at the same compute scale and achieves 2.5x model complexity growth.

The shift from per-surface to unified training

Before 2026, Meta trained separate models optimised for Facebook Feed, Instagram Stories, Reels and other surfaces. According to Meta, GEM shifts from isolated per-domain optimizations to a single central foundation model that learns from data across all of the company's ad platforms and now jointly predicts click-through rate across all surfaces and conversion rate across the major optimization goals.

That architectural change is what enables the claimed performance gains, not simply adding more compute. A unified model can learn from patterns that span surfaces (someone who watches Reels heavily but converts after seeing a Feed ad, for example) in a way that separate models could not.

Meta uses a post-training technique called Solaris that speculatively precomputes user-ad embeddings before a live ad request needs them, combined with a structured representation technique called LoopFM, which doubles intelligence transfer effectiveness beyond standard knowledge distillation at no additional compute cost or latency.

Training efficiency doubled while compute scaled 4x

Meta doubled end-to-end training efficiency to 20-25% Model FLOPs Utilization (MFU) while scaling training FLOPs 4x in 12 months, by co-designing kernels, precision, parallelism, networking, and memory together.

MFU measures how much of a GPU's theoretical compute capacity is actually used during training. Reaching 20-25% at this scale is notable because scaling to thousands of GPUs typically introduces communication overhead that degrades per-GPU efficiency. Meta achieved both higher throughput and higher efficiency simultaneously, which requires system-level optimisation beyond simply adding more hardware.

In January 2026, the company said it had doubled the GPU cluster used to train GEM and added a sequence learning architecture, producing a 3.5% lift in Facebook ad clicks and more than 1% gain in Instagram conversions. By the second quarter of 2026, CFO Susan Li credited GEM and user understanding models together with an 8.3% increase in ad clicks and a 15.7% uplift in conversions on Facebook.

The progression of reported figures shows GEM's impact scaling over time as Meta expanded the model and its training data. The 18.9% conversion figure represents cumulative gains over the past year, not a single update.

What this means for anyone managing Meta ads

The unification of GEM reduces the value of per-surface manual optimisation. If a single model now jointly predicts performance across Feed, Stories, Reels and off-site, then the advertiser's ability to express a preference for one surface over another becomes less relevant to the actual auction outcome.

That shift has been underway since Meta began consolidating campaign types under Advantage+ automation, but GEM's architecture makes it structural rather than optional. The system is designed to learn cross-surface patterns that a human media buyer cannot see or articulate in campaign settings.

This does not mean campaign structure is irrelevant. Creative quality, offer clarity, and conversion signal strength still determine whether the auction has anything worth optimising. What changes is the return on effort from manually allocating budget between placements or attempting to game delivery by isolating surfaces.

Our breakdown of Meta's Muse agent covered how AI intermediaries are reshaping the relationship between brands and audiences. GEM extends that dynamic into the ad auction itself. The advertiser's role shifts from directing delivery to supplying the inputs (creative, landing pages, conversion events) that the system can learn from.

The broader automation direction

Meta says it will extend GEM across all platform data, both organic and ads; build toward "LLM-scale long context"; scale GEM on larger clusters; and use AI-driven optimization to refine architecture alongside hardware.

The reference to organic data is significant. Currently, GEM trains on ad engagement and conversion signals. Extending training to include organic content interactions would give the model visibility into user behaviour that precedes any exposure to paid media, potentially improving cold-start performance for new campaigns and better identifying users in early consideration stages.

Meta will evolve GEM to reason with inference-time scaling to optimize compute allocation, power intent-centric user journeys, and enable agentic, insight-driven advertiser automation that drive higher ROAS.

"Inference-time scaling" refers to techniques borrowed from LLM serving, where the model can allocate more compute to harder predictions at the moment they are needed, rather than treating every ad auction with uniform computational effort. That would let Meta spend more compute on high-value auctions (someone likely to convert, high-margin products) and less on low-probability impressions, improving system-wide efficiency and advertiser outcomes simultaneously.

The roadmap Meta has laid out positions GEM as the foundation of a fully automated ads system where an advertiser provides a URL or product feed and the system handles creative generation, audience targeting, placement selection and budget allocation. That vision has been stated publicly since mid-2025, and GEM's architecture is explicitly designed to enable it.

For social media marketers, the near-term implication is that the shift toward AI-managed campaigns is not a feature toggle but an infrastructure reality. The auction is already running on a unified model optimised for outcomes Meta can measure (clicks, conversions, revenue) rather than for outcomes an advertiser can specify (reach this audience, on this surface, at this frequency). Aligning your measurement and creative strategy with what the system is actually optimising for is now table stakes, not advanced practice.

What to check in your own account

Meta's performance claims are system-wide aggregates. Individual account performance will vary based on vertical, creative quality, conversion signal strength, and budget scale. The questions worth asking are structural, not whether your account hit 18.9% improvement:

• Has your cost per result trended down over the past year while volume held or grew? That would be consistent with GEM's reported impact, though not proof of it.
• Are campaigns with broad targeting and multiple placements now outperforming tightly controlled setups that used to win? If so, that is the unified model doing work across surfaces that isolated campaigns cannot access.
• Are you still manually pausing placements or surfaces because they "look expensive" in the breakdown? If GEM is jointly optimising across surfaces, the per-surface cost might be high because that is where the conversions are, not because delivery is broken.

The shift to a unified foundation model does not make creative, offer, or funnel strategy less important. It makes those the only levers that still move the outcome in a direction you choose. Everything else is increasingly decided by a model trained on behaviour patterns across billions of users and millions of advertisers, optimised for an aggregate outcome you participate in but do not control.