The Pulse
Meta Scales GEM Ads Model Across Platforms and Thousands of GPUs
Meta says its GEM ads model now combines predictions across Facebook, Instagram, Reels and other advertising surfaces. The company reports a 4x increase in training compute and doubled end-to-end training efficiency on thousands of GPUs.
AI.info Team · Written with AI ·
One model for Meta’s ad surfaces
Meta says its updated Generative Ads Recommendation Model, or GEM, now works as a central system for ad ranking across Facebook, Instagram, Reels and other surfaces. In a post published October 13, the company describes a shift away from separate prediction systems for individual platforms and objectives toward one model trained on signals from across its advertising products.
The model jointly predicts clicks and conversions across surfaces and advertiser goals, including off-site conversions and mobile app installs. Meta says that engagement patterns learned in data-rich click tasks can help with conversion predictions, where signals are less common. The company says 3.6 billion people use at least one of its apps each day; it presents that scale as part of the challenge of combining varied user activity and ad data.
Meta reports an 18.9% improvement in conversions and a 12% improvement in click-through rate across its ad surfaces and objectives. It separately attributes conversion uplifts of 5.8% on Facebook, 4.1% on Instagram and 1.6% for off-site conversions among targeted users over the past year. The post does not explain the comparison baselines or measurement method for those figures.
Kunlun changes how GEM scales
The new generation uses Kunlun, an architecture Meta says is co-designed with recent GPU hardware. Compared with its earlier Interformer architecture, Kunlun delivers twice the performance at the same computation scaling, according to the post. Meta also says the latest GEM generation grew model complexity 2.5 times from earlier versions.
Kunlun combines attention, sequence summarization and cross-feature interactions, including a mixture-of-experts design. One change, block attention, replaces full self-attention with sparse attention arranged in blocks; Meta says this lowers sequence-length computation from quadratic to linear relative to the window size. Another, Component Skip Connection, selectively bypasses sequence modules and reduces FLOPs by 43%, with what Meta describes as negligible quality loss.
The architecture also supports longer user histories. Meta says it doubled the long-engagement context length in production and built the infrastructure to scale that context to 25 times its earlier level. The post presents the 25-times figure as a capability, not as the context length currently deployed.
GEM’s knowledge reaches smaller models
Meta does not run the large foundation model against every ad candidate during a live request. Instead, it describes LoopFM as a way to compress GEM’s representations and historical user context so that downstream models can learn from them. Meta says this approach doubles transfer effectiveness compared with the knowledge-distillation system used in the previous version, without additional compute cost or latency.
A second system, Solaris, precomputes representations for user-ad pairs that may be needed in a future request. The work runs asynchronously as ad ranking narrows the candidate set, then supplies the representation near the time a smaller model needs it. If an exact user-ad representation is unavailable, the system can use recent representations from the same user or behaviorally similar users, according to Meta.
Thousands of GPUs, with specialized training
GEM trains on several thousand of Meta’s latest-generation GPUs. Meta says it increased total training FLOPs fourfold over the past year while doubling end-to-end training efficiency to 20%–25% model FLOPs utilization, a measure of how effectively GPUs perform computation during training.
The training stack uses recommendation-specific kernels, including Jagged Flash Attention and Generalized Dot-Product Attention, alongside MXFP8 precision for attention and feed-forward layers. Meta also describes topology-aware five-dimensional parallelism: different strategies distribute dense and sparse model parameters across GPUs, with the cluster’s network layout informing how they communicate. The company’s post frames hardware, software and network design as a joint effort rather than a matter of adding more GPUs.
The reported gains come with limited measurement detail
Meta’s account ties the model’s scaling to measurable advertising outcomes, but the announcement does not provide the underlying evaluation design, comparison periods or traffic breakdowns behind its performance figures. That makes the reported conversion and click-through improvements difficult to compare with results from other systems using different baselines. What the post does specify is the engineering target: a shared ads model trained across thousands of GPUs, with 20%–25% end-to-end model FLOPs utilization after a year in which training compute grew fourfold.