Meta details 2026 GEM ads model: Kunlun architecture brings 2.5x complexity growth
Meta_Engineers · x · 2026-10-21
Meta's engineering team shared how it scaled the 2026 generation of GEM, the Generative Ads Model — the central model and compute engine behind ads across Meta. Introduced last year as the industry's largest recommendation foundation model, GEM now pushes further on architecture, cross-surface cross-domain learning, intelligence transfer, and LLM-scale infrastructure.
Key points:
- Kunlun architecture: a new LLM-style architecture co-designed with the latest GPUs, delivering 2.5x model complexity growth over earlier GEM versions and 2x longer per-person engagement history, while modeling recommendation-specific massive feature interactions.
- Ecosystem-wide learning: GEM is the first model to jointly predict clicks and conversions across Meta's ads surfaces.
More from Companies & People
- Ex-OpenAI research chief Bob McGrew raises $230M for Arda at $1.5B valuation — Kr00ney · 2026-10-21
- Tyler and Marc's public beef is the best entertainment on X right now — jdluk87 · 2026-10-21
- Pivotal opens London research fellowship: 15 weeks, £15-18k stipend, mentors from Anthropic and DeepMind — davidmanheim · 2026-10-21
- Linear's founder: agent-built products become feature piles unless you have a point of view — wateriscoding · 2026-10-21
- LatePost: How twentysomething researchers — including a 17-year-old — shaped Moonshot's K3 — dotey · 2026-10-21
- PyTorch Conference 2026 opens in San Jose; Nebius and Nscale join as Platinum members — PyTorch · 2026-10-21