---
title: "Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI"
description: "Meta introduced GEM, the industry’s largest recommendation system foundation model, and its 2026 generation shifts from isolated per-domain optimizations on Facebook, Instagram, and Reels to a single central foundation model. The Kunlun architecture, co-designed with the latest GPU hardware, enables 2.5x model complexity and demonstrates LLM-like scaling laws with block-aligned sparse attention reducing sequence complexity. The LoopFM post-training technique transfers GEM’s internal representations to downstream models, doubling intelligence transfer at no additional compute cost or latency. GEM delivered 18.9% conversion improvement and 12% ad click-through rate improvement across all Meta ad surfaces and funnel stages, including 5.8% conversion uplift on Facebook, 4.1% on Instagram, and 1.6% on off-site conversion. Training efficiency doubled to 20-25% model FLOPS utilization while scaling total training FLOPs by 4x across thousands of GPUs using custom kernels and 5D parallelism."
---

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

[Meta](https://yomu.fyi/company/meta) · Huayu Li · Oct 13, 2026

**Type:** Announcement

## Summary

Meta introduced GEM, the industry’s largest recommendation system foundation model, and its 2026 generation shifts from isolated per-domain optimizations on Facebook, Instagram, and Reels to a single central foundation model. The Kunlun architecture, co-designed with the latest GPU hardware, enables 2.5x model complexity and demonstrates LLM-like scaling laws with block-aligned sparse attention reducing sequence complexity. The LoopFM post-training technique transfers GEM’s internal representations to downstream models, doubling intelligence transfer at no additional compute cost or latency. GEM delivered 18.9% conversion improvement and 12% ad click-through rate improvement across all Meta ad surfaces and funnel stages, including 5.8% conversion uplift on Facebook, 4.1% on Instagram, and 1.6% on off-site conversion. Training efficiency doubled to 20-25% model FLOPS utilization while scaling total training FLOPs by 4x across thousands of GPUs using custom kernels and 5D parallelism.

## Context

Meta operates dozens of ad surfaces including Facebook, Instagram, and Reels, each previously with independent prediction systems. Scaling a single foundation model across these heterogeneous surfaces and objectives required new architecture and training techniques.

## Approach / What changed

Built the Kunlun LLM-style architecture co-designed with GPU hardware, enabling 2.5x model complexity. Used LoopFM for efficient intelligence transfer to downstream models. Created a single unified compute engine for cross-surface, cross-domain learning. Developed custom kernels and 5D parallelism to double training efficiency to 20-25% MFU while scaling FLOPs 4x.

## Takeaways

- GEM’s Kunlun architecture uses block-aligned sparse attention to reduce sequence complexity from quadratic to linear relative to window size, enabling longer user engagement contexts.
- The LoopFM post-training technique transfers GEM’s internal representations to downstream models, doubling intelligence transfer at no additional compute cost or latency through computation sharing.
- GEM achieved 18.9% conversion improvement and 12% CTR improvement across Meta’s ad surfaces, with specific uplifts of 5.8% on Facebook, 4.1% on Instagram, and 1.6% on off-site conversion for targeted users.

## FAQ

### How does GEM handle the latency challenge of serving a large model in real time?

GEM uses LoopFM, a post-training technique that transfers internal representations to downstream models through computation sharing, doubling intelligence transfer at no additional compute cost or latency. It also uses elastic pre-computation (Solaris) to harvest gains from the upstream model.

### What architecture innovations enable GEM's scaling?

GEM uses the Kunlun architecture with block-aligned sparse attention, mixture-of-experts cross-interaction, and deep co-design with GPU hardware to enable 2.5x model complexity growth and LLM-like scaling laws.

### What performance improvements did GEM deliver?

GEM delivered 18.9% conversion improvement and 12% ad click-through rate improvement across Meta's ad surfaces, including 5.8% conversion uplift on Facebook, 4.1% on Instagram, and 1.6% on off-site conversion for targeted users over the past year.

**Tags:** [AI](https://yomu.fyi/topic/ai), [GPU](https://yomu.fyi/topic/gpu), [LLMs](https://yomu.fyi/topic/llm), [Meta](https://yomu.fyi/topic/meta), [Recommendation Systems](https://yomu.fyi/topic/recommendation-systems)

- Source: [Meta](https://engineering.fb.com/2026/10/13/ml-applications/generative-ads-model-gem-scaling-frontier-ads-ai)
- Source URL: https://engineering.fb.com/2026/10/13/ml-applications/generative-ads-model-gem-scaling-frontier-ads-ai
- Ingested by Yomu: 2026-10-13T18:00:13.046Z

[Read original post](https://engineering.fb.com/2026/10/13/ml-applications/generative-ads-model-gem-scaling-frontier-ads-ai)