The lift across ad surfaces is Meta's own account
Meta's engineering post says the 2026 generation of GEM stops optimizing Facebook, Instagram, and Reels as separate systems and uses one central foundation model that learns from data and signals across its ads platforms. They report a 18.9% conversion improvement and a 12% click-through improvement across ad surfaces and objectives. Over the past year, targeted users saw an additional 5.8%, 4.1%, and 1.6% conversion uplift on Facebook, Instagram, and off-site conversion. Those are Meta's online results. The post does not include an outside retest.
They also say 3.6 billion people use at least one Meta app each day. That is Meta's user figure, not the sample size of this comparison.
[1]
At the same compute, Kunlun is written as twice the previous version
Against the previous version, Interformer, Kunlun delivers 2x performance at the same computation scaling. At their largest pre-training scale they report 2.5x model complexity growth from earlier versions of GEM. In production, long engagement context grew by 2x. They also say the infrastructure can scale long context to 25x. That 25x is a capability claim, not a length already rolled out everywhere.
A mechanism that skips selected sequence modules reduces FLOPS by 43%, which they call a negligible quality loss. After gathering a class of dot-product work into one kernel, forward speed is up to 2x. LoopFM passes the central model's representations and chronological history to downstream models. With computation sharing, intelligence transfer doubles relative to the previous distillation, at what they say is no extra compute or latency.
[1]Training FLOPs are 4x, and utilization is written in the low twenties
End-to-end training efficiency doubled over the past 12 months, to 20–25% model FLOPS utilization, with 4x total training FLOPs, on thousands of the latest GPUs. The post describes this as training at the meeting point of recommendation models and language-model scale. It does not include an outside bill to compare against.
Clicks and conversions used to have a separate system for each surface and objective. GEM is described as the first model that jointly predicts click-through rate across surfaces and conversion rate across the main objectives. That "first" is their wording.
[1]