Meta yesterday published a technical account of the 2026 generation of its Generative Ads Model, known as GEM, claiming the system delivered an 18.9% conversion improvement and a 12% click-through rate improvement across its ad surfaces and objectives. The engineering post, credited to seven Meta engineers, describes how the company replaced separate prediction systems for Facebook, Instagram and Reels with one central foundation model, and how it moves that model's output into the hundreds of smaller models that actually serve ads in real time.

In Short

Meta built one very large AI model that learns from what people do across Facebook, Instagram and its other apps, and uses it to teach the smaller models that pick which ad you see. Meta says this made people click on ads 12% more and convert 18.9% more, which matters to anyone paying for ads on those apps. The figures come only from Meta, with no baseline or method attached, so they describe what Meta says changed rather than something an advertiser can confirm in their own account.

What Meta published

The post appeared today on Engineering at Meta, the company's technical blog, under the ML Applications category. Its byline lists Huayu Li, Darren Liu, Xiaoyi (Leo) Liu, Ellie Wen, Jackie (Jiaqi) Xu, Chunzhi Yang and Reva Srinivasan, followed by an acknowledgements section naming more than 130 contributors. It is the third long-form GEM document Meta has released in under a year. The first, dated November 10, 2025, described GEM as "The Central Brain Accelerating Ads Recommendation AI Innovation." A second, dated August 3, 2026, covered training efficiency. Today's post consolidates and updates both.

Meta opens with a large claim. According to the company, GEM is "the industry's largest recommendation system (RecSys) foundation model," and it "established Meta's first LLM-inspired paradigm for ads recommendation." Meta gives no parameter count, no dataset size and no comparison against named systems from other companies, so the claim cannot be tested from the document.

The central change in the 2026 version is structural. According to Meta, GEM "shifts from isolated per-domain optimizations - such as Facebook, Instagram, Reels - to a single central foundation model." That model learns from data across all of the company's ad platforms. Meta says it now jointly predicts click-through rate across all surfaces and conversion rate across the major optimization goals, from off-site conversions to mobile app installs. "Previously, each surface and each prediction objective had its own system," the post states.

The performance figures, and what is missing from them

The headline numbers sit in a single paragraph. According to Meta, "GEM has delivered 18.9% conversion improvement and 12% ad click-through rate (CTR) improvement across Meta's ad surfaces and objectives." The post adds that GEM "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."

Several details are absent. The post does not say what the 18.9% and 12% figures are measured against, over what period they accrued, whether they come from A/B tests or offline evaluation, or whether they are averages weighted by spend, by impressions or by advertiser. The second set of figures applies to "targeted users," a phrase the post does not define. It is also unclear whether the 5.8%, 4.1% and 1.6% sit inside the 18.9% or on top of it; the word "additionally" suggests the latter, but the post does not reconcile the two sets.

These are vendor-supplied figures about Meta's own system. They are not independently audited.

How the numbers compare with earnings disclosures

Meta has been citing GEM in quarterly results for more than a year, and those disclosures do not map neatly onto today's figures. In July 2025, Mark Zuckerberg said the new ads recommendation model had driven roughly 5% more ad conversions on Instagram and 3% on Facebook in the second quarter of 2025. In October 2025, Meta said improvements to GEM's architecture doubled the performance benefit from adding data and compute. 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.

The figures then grew. In April 2026, Meta attributed a more than 6% increase in landing page view ad conversion rates to Lattice enhancements combined with GEM architecture advances. 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.

Today's post gives Facebook a 5.8% conversion uplift "for targeted users." The July earnings call gave Facebook 15.7% for a combination of GEM and user understanding. The two numbers may describe different populations, different periods or different attribution between models, but the post does not explain the gap. The 1.6% off-site conversion figure also matches, to the decimal, the lift Meta reported in April for expanding its Adaptive Ranking Model to off-site conversions. Whether this is the same measurement credited to two systems or a coincidence is not stated.

Kunlun: rebuilding the architecture for scale

Most of the post concerns how Meta made a recommendation model behave more like a large language model when compute is added. The premise is stated plainly. "Empirical scaling laws demonstrate that model performance predictably improves with compute, parameter, and data size using transformer model architecture," Meta writes. "However, recommender system models do not exhibit similar scaling properties with transformer due to massive heterogeneous features and shifting distributions."

Ad ranking data is messy in ways text is not. A single prediction mixes user demographics, ad content representations and long sequences of past events, each event carrying its own feature set. Those features shift constantly as campaigns start and stop. Meta's answer is a new architecture called Kunlun, which replaces the previous design, Interformer. According to Meta, Kunlun "delivers 2x performance gains with the same computation scaling" compared with Interformer, and enabled 2.5x model complexity growth over earlier versions of GEM.

Each Kunlun layer has three parts, according to the post. An attention component, with customised attention and feed-forward networks, captures long dependencies in user behaviour. A summarization component extracts what Meta calls "critical user insights" from sequences and bridges sequence and non-sequence data. A cross-interaction component, expanded to a mixture-of-experts design, models interactions between the two feature types at scale.

Engineering the cost down

The architecture was co-designed with what Meta describes only as "the latest GPU hardware." The post does not name a chip vendor or say whether Meta's own MTIA accelerators are involved in GEM training. Three cost-reduction techniques are described.

First, block attention. Replacing full self-attention with block-aligned sparse attention reduces the cost of sequence length "from quadratic to linear relative to window size," according to Meta.

Second, a reformulation called Generalized Dot-Product Attention, or GDPA. Meta had earlier added Per-Example Feed-Forward Networks (pFFN) to adapt the standard feed-forward layer to each example's distribution. GDPA rewrites that component as attention operations so that many attention-like operations can run in a single GPU kernel, which Meta says delivers "up to 2x forward speedup."

Third, expert parallelism for the cross-interaction component, which lets each GPU process a subset of expert sub-networks in parallel.

Two more mechanisms target sequence length directly. CompSkip, a component skip connection, selectively deactivates sequence modules in some layers and routes outputs through residual connections. Meta says this cuts FLOPs by 43% "with negligible quality loss," an outcome resting on what the post itself calls an "assumption that the network can recover quickly after re-activating." OneSeq unifies different user sequence sources into one chronological stream after feature aggregation. The post gives an example: an event in a purchase sequence is described by ad ID and account ID, while an event in a post sequence is described by actors and object ID. Those cannot be merged at the data source, so OneSeq merges them later, which Meta says maximises GPU efficiency.

Longer histories, bigger models

The practical result is longer user histories. According to Meta, GEM "scaled the long engagement context length by 2x in production," and the infrastructure can scale long context to 25x. The 25x figure is a capability, not a production setting.

Meta reports two findings from this work. Prediction quality follows a power law with sequence length, and larger models benefit disproportionately from longer sequences. The post frames this as a "virtuous cycle" in which more compute justifies longer sequences and longer sequences justify more compute. One caveat sits in a figure caption: the scaling law pattern shown was measured "on the light-weight model of GEM with small parameter size," not the production model. The post separately says Meta "observed an LLM-like scaling law at our largest pre-training scale" but publishes no chart or numbers for that observation.

For advertisers, the detail matters because it describes how far back Meta's ranking looks into a person's behaviour. Meta does not say what the baseline sequence length was, so "2x" cannot be converted into days, events or sessions.

Post-training: LoopFM and Solaris

A model of this size cannot be run inside a live ad auction. Meta acknowledges this directly. "Serving a model at this scale in real time is prohibitively expensive and drives up latency, compromising both user experience and advertiser ROI," the post states. GEM therefore works as a teacher. Its value, according to Meta, "depends on how effectively its intelligence transfers to the hundreds of vertical models that serve ads in real time."

The previous version relied on knowledge distillation, where downstream models learn from the foundation model's soft labels. Meta calls that approach "effective but limited in bandwidth," because downstream models only receive a compressed view of GEM at each training step.

LoopFM

The 2026 version adds what Meta calls a structured representation, with LoopFM as the named example. The process has three steps. GEM's output from important layers is compressed into compact representations describing each event it processed. Those representations are then arranged with the representations GEM previously generated for the same user into a sequential context window, which Meta says captures "user intent dynamics across all surfaces and traffic." Downstream models then learn from that structured history during both training and inference.

According to Meta, LoopFM doubles intelligence transfer together with the existing knowledge distillation "at no additional compute cost or latency," because the representation generation shares computation with the soft-label generation already done for distillation. The post claims "2x of transfer effectiveness beyond knowledge distillation" but does not define how transfer effectiveness is measured.

Solaris

The second mechanism, Solaris, handles freshness. "Evaluating the large-scale model inside a live ad request is not affordable," Meta writes, so Solaris moves the computation earlier in time. It borrows the idea of speculative decoding from LLM serving: as multi-stage ranking narrows a live request to its top ad candidates, Solaris evaluates those user-ad pairs with the foundation model in asynchronous background jobs, then delivers the embeddings "just-in-time on the serving hot path."

The result, according to Meta, is that vertical models receive near-real-time knowledge of how a specific user is likely to act on a specific ad candidate, without the latency cost of running GEM inline. The background computation runs on elastic capacity with no latency budget. Meta published a separate post on October 12, 2026, titled "Turning Opportunistic AI Inference Capacity Into Better Ads Recommendations," which appears in the same ML Applications listing, though today's post does not link the two explicitly.

The fallback behaviour is worth noting. When an exact user-ad embedding is unavailable, "the system degrades rather than drops," according to the post. It substitutes an aggregate of the user's own recent embeddings across the same ad groups, or "borrows from behaviorally similar users." In practice, a person's predicted response to an ad can be partly inferred from how other people with similar behaviour responded, a process that sits entirely inside Meta's systems and is not exposed to advertisers.

One model across every surface

The cross-surface training section explains how Meta merged data that was previously siloed. Meta trained GEM on what it describes as "all-surface data (e.g., Facebook and Instagram Feed and Reels/video)" together with deep-funnel optimization data such as off-site conversion and mobile app installs. The post cites 3.6 billion people using at least one Meta app each day, consistent with the 3.60 billion family daily active people Meta reported for June 2026.

The logic is about data density. Clicks are plentiful; conversions are scarce. According to Meta, "the generalizable representations learned on data-rich click tasks transfer into data-sparse conversion tasks." Engagement patterns on one surface can now inform predictions of purchase behaviour on another.

Two techniques make that possible, according to the post. A cost-free data mixing strategy loads data virtually from different combinations of pipelines without physically merging them, avoiding additional storage costs. Meta also controls the mixture proportions during offline and online training to stay close to the real order in which data arrives, which it says mitigates instability when switching domains and "overmemorization for the recent domain data." A domain expert framework then pairs a shared expert, which learns a unified representation across all domains, with domain-specific experts that capture patterns particular to one surface. Both are built on Kunlun.

Meta frames the outcome as "simultaneous ad performance gains across every surface and funnel stage." The post presents this as the first such training recipe for large-scale cross-surface, cross-domain unification in recommendation systems, a priority claim it does not support with references.

It is worth distinguishing this from consolidation of the serving stack. GEM is a single central model, but the post is explicit that hundreds of vertical models still serve ads, and that downstream models "still train with the specific surface or traffic data." Meta's separate consolidation effort, Lattice, had by October 2025 reduced the number of ads ranking and recommendation models by about 100 since 2023. Today's post does not mention Lattice by name.

Training at LLM scale: 20-25% MFU

The final technical section restates material from Meta's August 3, 2026 companion post on training efficiency. GEM now trains "on several thousand of the latest-generation GPUs," according to Meta. No exact number is given. Earlier this year, Meta said it had doubled the GPU cluster size for training GEM in the fourth quarter of 2025.

Meta measures training efficiency through end-to-end model FLOPS utilization, which it decomposes as local MFU multiplied by scaling ratio. Local MFU measures how close a single GPU runs to its hardware roofline; scaling ratio measures how much of that single-GPU performance survives distribution across thousands of devices. According to the post, GEM's end-to-end efficiency doubled to 20-25% MFU over the past 12 months while total training FLOPs grew 4x. The post does not give the starting MFU, though "doubled" implies roughly 10-12.5%.

The figure needs context. An MFU of 20-25% means three quarters or more of the theoretical GPU compute is not being used for model arithmetic. Meta's own explanation is that "today's data center GPUs and their software stacks are mostly optimized for LLM workloads, whereas recommendation workloads have a fundamentally different profile." Recommendation models handle sparse, jagged inputs and huge embedding tables, which are harder to keep a GPU busy with.

The techniques cited fall into two groups. For compute efficiency, Meta built a custom kernel library including Jagged Flash Attention (JFA), GDPA and block attention, and uses mixed ultra-low precision training, including MXFP8 for attention and MLP layers. For scaling efficiency, Meta uses what it calls topology-aware 5D parallelism with collectives that do not occupy Streaming Multiprocessors: 2D FSDP plus expert parallelism for dense parameters, combined with fully sharded 2D model parallelism for sparse parameters, all designed around Meta's multi-tiered network.

"Training a foundation model at the intersection of LLMs and recommendation systems is a co-design problem, not a software problem or a hardware problem alone," according to the post. The engineering detail is unusually specific for Meta. The commercial detail is not: there is no figure for cost per training run, energy use or the share of Meta's capital budget allocated to GEM.

Why this matters for advertisers

Meta's 2026 capital expenditure guidance stands at $130 billion to $145 billion, narrowed in July, and its infrastructure chief has said Meta will spend well over $100 billion this year on AI infrastructure. GEM is one of the few places where Meta ties that spending to a measurable ad outcome. Today's post functions partly as a public justification of that link.

For media buyers, three points stand out.

The first is that cross-surface learning changes what an ad set optimizes on. When conversion prediction for an Instagram campaign draws on click behaviour from Facebook Feed, the separation between placements becomes weaker inside the model, even where campaign reporting still shows them separately. Advertisers with low conversion volumes may benefit most from click-to-conversion transfer, if the claims hold, because their own conversion data is the sparsest.

The second is opacity. None of the mechanisms described - Kunlun, LoopFM, Solaris, the domain expert framework - has a setting in Ads Manager. Their effects arrive through Advantage+ and Meta's automated delivery, which advertisers have questioned before. PPC Land reported in November 2025 on advertiser skepticism about Meta's AI automation, including one incrementality test in which Advantage+ produced only 17% of the conversions Meta's attribution reported. A 12% gain in click-through rate on Meta's own measurement says nothing about incremental sales on any given account.

The third is data scope. Meta writes that it intends to "extend GEM to learn from all of Meta's platform data, capturing behavior across organic and ads engagements spanning text, images, and video." That direction follows Meta's decision, effective December 16, 2025, to use conversations with Meta AI for ad and content recommendations, excluding sensitive categories such as health, religion and sexual orientation. Today's post does not mention Meta AI chat data, regional restrictions or the European Union, and it does not say whether GEM's cross-surface training covers users in jurisdictions where Meta offers less personalized ads.

A wider trend in recommendation systems

Meta is not the only platform publishing this kind of material. In March, LinkedIn described how it rebuilt its feed with LLMs and GPU-powered ranking, processing more than 1,000 historical member interactions with transformer attention. The shared direction is clear: borrow LLM architecture, scale context length, and accept higher training costs in return for prediction gains.

Meta's July earnings also brought a separate system, the Meta Generative Recommender, which uses LLMs to reason about ad content and user preferences together and select one best match per person. Today's post does not mention it, and the relationship between that system and GEM is not explained.

What comes next

Meta lists four directions. The company 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. It also plans to "evolve the post-training paradigm with reasoning capability" to transfer more of GEM's output to downstream models.

"This will unlock better ad relevance and help advertisers connect with the right audiences more efficiently," according to Meta. No dates, milestones or target figures accompany any of the four directions.

Timeline

Summary

Who: Meta, through an engineering post credited to Huayu Li, Darren Liu, Xiaoyi (Leo) Liu, Ellie Wen, Jackie (Jiaqi) Xu, Chunzhi Yang and Reva Srinivasan. The changes affect advertisers buying across Facebook, Instagram, Reels and Meta's other surfaces.

What: A technical account of the 2026 generation of GEM, Meta's central ads recommendation foundation model. Meta claims an 18.9% conversion improvement and 12% CTR improvement, and describes a new architecture (Kunlun), two post-training transfer mechanisms (LoopFM and Solaris), a cross-surface training recipe and 20-25% MFU training efficiency with 4x more training FLOPs. No baseline, period or methodology is given for the performance figures.

When: Published October 13, 2026, following earlier GEM posts on November 10, 2025 and August 3, 2026.

Where: Engineering at Meta, the company's technical blog. The model trains on several thousand GPUs in Meta's data centers and informs ad delivery across all Meta ad surfaces globally; the post does not address regional differences.

Why: Meta is replacing separate per-surface prediction systems with one model so that signals from data-rich click tasks improve data-sparse conversion predictions, and it is using GEM to justify part of a 2026 capital budget of $130 billion to $145 billion. For advertisers, the changes alter how delivery decisions are made without adding any controls or independent measurement.