The thesis is that Meta is treating cross-surface recommendation as a shared model-and-compute problem, with specialized downstream models consuming transferred representations rather than each domain remaining an isolated optimization target.
The mechanism is a central model that jointly predicts clicks and conversions, paired with post-training and infrastructure intended to make its representations usable downstream. A central model jointly predicts clicks and conversions, while post-training and infrastructure transfer its representations to downstream models; Meta says LoopFM doubles intelligence transfer without additional compute cost or latency. Meta says the new generation is 2.5 times more complex than earlier GEM versions and reports four times the total training FLOPs while doubling end-to-end training efficiency to 20–25% model FLOPs utilization. It also says LoopFM doubles intelligence transfer without additional compute cost or latency. These are observed facts in Meta’s engineering account, not independently verified measurements. The reported outcome is an 18.9% conversion improvement and 12% CTR improvement across surfaces and objectives.
Why now: Meta is presenting the next step after its earlier GEM system as a scaling strategy across dozens of surfaces and advertiser goals, rather than a set of local model upgrades. The first implication is that shared representations may let data-rich click tasks improve data-sparser conversion tasks across domains. Shared representations may let data-rich click tasks improve data-sparser conversion tasks across domains. The evidence boundary is that performance, efficiency, and causal attribution figures come from Meta; the supplied source does not establish independent replication or disclose a full evaluation methodology. Performance, efficiency, and causal attribution figures come from Meta; the supplied source does not establish independent replication or disclose a full evaluation methodology. Watch whether Meta publishes comparable results or operational details that clarify how gains vary by surface and objective. The new branch records a distinct architectural shift to central cross-surface ads modeling; no prior New Runtime record was supplied.