Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI

Huayu Li, Darren Liu, Xiaoyi (Leo) Liu, Ellie Wen, Jackie (Jiaqi) Xu, Chunzhi Yang, Reva Srinivasan13 Oct 2026

Problem

Meta shows ads on dozens of surfaces (Facebook, Instagram, Reels, Stories) to 3.6 billion people a day. Each surface and objective had its own prediction system, so learning did not carry over. Ranking models also did not follow transformer scaling laws, because features are heterogeneous and distributions shift. A big central model would help, but it is too expensive to run inside every live ad request.

Example from the article: the 2026 generation of the Generative Ads Model (GEM) replaces per-surface systems with one central model that predicts clicks and conversions for all surfaces.

Solution in brief

Build one large model with an architecture co-designed with GPU hardware (Kunlun), so it scales with compute and user-history length. Train it jointly on all surfaces and objectives with a domain-expert layout. Move its knowledge into the production models offline (LoopFM representations) and just before serving (Solaris precomputation), never inside the request.

Questions this case answers
  • How to scale a large ranking model with compute and user-history length when heterogeneous features break the usual scaling behavior
  • How to transfer the predictions of a model too expensive for real-time serving into many lightweight production models without adding latency
  • How to train one shared model across several domains and objectives where some objectives have far fewer labels
  • How to make a recommendation ranking model scale with compute and sequence length the way LLMs do?

Tools and methods

  • Kunlun — LLM-style ranking architecture for long sequences with mixture-of-experts cross interaction.
  • Block attention, CompSkip, OneSeq — cheaper long sequences, unified event sources.
  • LoopFM — precomputed compact representations of GEM plus the user's earlier representations, fed to downstream models.
  • Solaris — speculative background precomputation of user-ad embeddings for likely candidates.
  • Domain experts and virtual data mixing — shared plus per-domain experts, no physical data merge.
Continue reading — it's free

Google or a one-time email code. No password.