Meta has unveiled the 2026 generation of its Generative Ads Model, GEM, which replaces separate models for each platform with a single central model covering all of the company's advertising. Introduced last year as the largest recommendation foundation model in the industry, GEM now aims to scale compute and knowledge transfer on next-generation hardware.
According to Meta, GEM delivered an 18.9% improvement in conversions and a 12% gain in ad click-through rate across its ad surfaces and objectives. Over the past year it also contributed conversion uplifts of 5.8% on Facebook, 4.1% on Instagram and 1.6% for off-site conversions among targeted users.
An architecture that breaks the scaling barrier for recommenders
Recommender models, unlike the transformers behind LLMs, have so far not shown predictable gains as compute, parameters and data grow, because of heterogeneous features and shifting distributions. To address this, Meta built an LLM-style architecture called Kunlun, co-designed with the latest GPUs to make larger, deeper models cost-efficient and to learn from long user engagement histories.
An arXiv paper describing Kunlun reports that it raises model FLOPs utilization from 17% to 37% on NVIDIA B200 GPUs and doubles scaling efficiency over prior methods. Meta says Kunlun delivers 2x the performance of its previous architecture, Interformer, at the same compute.
Moving intelligence into production models
Serving a model of this scale in real time would be prohibitively expensive and slow. GEM therefore transfers its internal representations and historical memory to user-facing models through a post-training technique. Meta says that, through computation sharing, the transfer doubles the gains of the knowledge distillation used in the previous version, at no extra compute cost or latency.
One engine for every surface
Meta's apps reach 3.6 billion people a day across dozens of surfaces and objectives, each of which previously ran its own prediction system. The central model jointly predicts clicks and conversions across all of them, from off-site conversions to mobile app installs, transferring signals learned on data-rich click tasks to data-sparse conversion tasks.
Training across thousands of GPUs
Training such a model across thousands of GPUs requires careful orchestration of compute, parallelism, communication and memory. Meta developed custom low-precision kernels and new 5D parallelism strategies, doubling end-to-end efficiency to 20-25% MFU with 4x total training FLOPs. Data-center GPUs and their software stacks are mostly optimized for LLM workloads, whereas recommendation workloads have a very different profile.
No comments yet. Be the first!