--- title: Meta’s Generative Ads Model (GEM): Scaling the Central... description: Meta details the 2026 generation of its Generative Ads Model (GEM), a unified foundation model replacing per-surface ad prediction systems across Facebook,... canonical: https://daily.dev/posts/meta-s-generative-ads-model-gem-scaling-the-central-model-and-compute-engine-for-frontier-ads-ai-qkcz7rbzd twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI | daily.dev og:description: Meta details the 2026 generation of its Generative Ads Model (GEM), a unified foundation model replacing per-surface ad prediction systems across Facebook,... og:url: https://daily.dev/posts/meta-s-generative-ads-model-gem-scaling-the-central-model-and-compute-engine-for-frontier-ads-ai-qkcz7rbzd og:image: https://api.daily.dev/og/posts/qkCz7RBzD.png og:image:alt: Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI **[Facebook Engineering ](https://daily.dev/sources/facebook_code)** · 16 min read · 0 upvotes · 0 comments ## Summary Meta details the 2026 generation of its Generative Ads Model (GEM), a unified foundation model replacing per-surface ad prediction systems across Facebook, Instagram, and Reels. The new Kunlun architecture enables 2.5x model complexity growth with LLM-like scaling laws, while post-training techniques (LoopFM, Solaris) double intelligence transfer to downstream models without added latency. A new training recipe supports cross-surface, cross-domain learning, and infrastructure improvements doubled end-to-end training efficiency to 20-25% MFU while scaling total training FLOPs 4x. Reported gains include 18.9% conversion improvement and 12% CTR improvement across Meta's ad surfaces over the past year. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What architecture does Meta's GEM ads model use to achieve LLM-like scaling for recommendation systems? Meta's GEM uses a custom LLM-style architecture called Kunlun, co-designed with the latest GPU hardware, which delivers 2x performance gains over the prior Interformer architecture at the same compute scale. It achieves 2.5x model complexity growth and enables long-context learning over user engagement history through block attention, generalized dot-product attention, and mixture-of-experts style cross-interaction. _Engineers scaling recommendation systems can track architecture innovations like Kunlun through daily.dev._ ### How does Meta transfer knowledge from a large foundation model to real-time ad serving models without added latency? Meta uses a post-training technique called Solaris that speculatively precomputes user-ad embeddings before a live ad request needs them, borrowing the idea of speculative decoding from LLM serving. Combined with a structured representation technique called LoopFM, this doubles intelligence transfer effectiveness beyond standard knowledge distillation at no additional compute cost or latency. _Teams balancing model intelligence against serving latency can follow techniques like this on daily.dev._ ### What GPU training efficiency did Meta achieve for its GEM ads recommendation foundation model? Meta doubled end-to-end training efficiency to 20-25% model FLOPS utilization (MFU) while scaling total training FLOPs by 4x over twelve months. This came from custom low-precision kernels (including MXFP8 attention and MLP), new 5D parallelism strategies, and topology-aware networking co-designed together across thousands of GPUs. _Developers optimizing GPU training throughput at scale can follow infrastructure deep dives like this on daily.dev._ ## Similar posts on daily.dev - [Meta’s Generative Ads Model \(GEM\): The Central Brain Accelerating Ads Recommendation AI Innovation](https://daily.dev/posts/meta-s-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation-jduz74tbf) · Facebook Engineering · 1 upvotes · 0 comments - [How Meta Built a New AI-Powered Ads Model for 5% Better Conversions](https://daily.dev/posts/how-meta-built-a-new-ai-powered-ads-model-for-5-better-conversions-plmz0bqth) · ByteByteGo · 0 upvotes · 0 comments - [Meta Details GEM Ads Model Using LLM-Scale Training, Hybrid Parallelism, and Knowledge Transfer](https://daily.dev/posts/meta-details-gem-ads-model-using-llm-scale-training-hybrid-parallelism-and-knowledge-transfer-21ryza9tt) · InfoQ · 0 upvotes · 0 comments - [GEM Training: How Meta Doubled the Efficiency of Its LLM-Scale Ads Foundation Model](https://daily.dev/posts/gem-training-how-meta-doubled-the-efficiency-of-its-llm-scale-ads-foundation-model-yuawarr3a) · Facebook Engineering · 2 upvotes · 0 comments - [Welcome Gemma 4: Frontier multimodal intelligence on device](https://daily.dev/posts/welcome-gemma-4-frontier-multimodal-intelligence-on-device-mkjiapoth) · Hugging Face · 8 upvotes · 3 comments --- Tags: [#recommendation-systems](https://daily.dev/tags/recommendation-systems) [View this post on daily.dev](https://daily.dev/posts/meta-s-generative-ads-model-gem-scaling-the-central-model-and-compute-engine-for-frontier-ads-ai-qkcz7rbzd) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI","url":"https://daily.dev/posts/meta-s-generative-ads-model-gem-scaling-the-central-model-and-compute-engine-for-frontier-ads-ai-qkcz7rbzd","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/meta-s-generative-ads-model-gem-scaling-the-central-model-and-compute-engine-for-frontier-ads-ai-qkcz7rbzd"},"datePublished":"2026-10-13T16:01:10.381Z","dateModified":"2026-10-13T16:41:19.666Z","description":"Meta details the 2026 generation of its Generative Ads Model (GEM), a unified foundation model replacing per-surface ad prediction systems across Facebook,...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/868d2408f5a6efff851a3707542a56c5?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/868d2408f5a6efff851a3707542a56c5?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Facebook Engineering\n","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Facebook Engineering\n","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/facebook","url":"https://daily.dev/sources/facebook_code"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/meta-s-generative-ads-model-gem-scaling-the-central-model-and-compute-engine-for-frontier-ads-ai-qkcz7rbzd","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"recommendation-systems","timeRequired":"PT16M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Facebook Engineering\n","item":"https://daily.dev/sources/facebook_code"},{"@type":"ListItem","position":3,"name":"Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/meta-s-generative-ads-model-gem-scaling-the-central-model-and-compute-engine-for-frontier-ads-ai-qkcz7rbzd#faq","mainEntity":[{"@type":"Question","name":"What architecture does Meta's GEM ads model use to achieve LLM-like scaling for recommendation systems?","acceptedAnswer":{"@type":"Answer","text":"Meta's GEM uses a custom LLM-style architecture called Kunlun, co-designed with the latest GPU hardware, which delivers 2x performance gains over the prior Interformer architecture at the same compute scale. It achieves 2.5x model complexity growth and enables long-context learning over user engagement history through block attention, generalized dot-product attention, and mixture-of-experts style cross-interaction. Engineers scaling recommendation systems can track architecture innovations like Kunlun through daily.dev."}},{"@type":"Question","name":"How does Meta transfer knowledge from a large foundation model to real-time ad serving models without added latency?","acceptedAnswer":{"@type":"Answer","text":"Meta uses a post-training technique called Solaris that speculatively precomputes user-ad embeddings before a live ad request needs them, borrowing the idea of speculative decoding from LLM serving. Combined with a structured representation technique called LoopFM, this doubles intelligence transfer effectiveness beyond standard knowledge distillation at no additional compute cost or latency. Teams balancing model intelligence against serving latency can follow techniques like this on daily.dev."}},{"@type":"Question","name":"What GPU training efficiency did Meta achieve for its GEM ads recommendation foundation model?","acceptedAnswer":{"@type":"Answer","text":"Meta doubled end-to-end training efficiency to 20-25% model FLOPS utilization (MFU) while scaling total training FLOPs by 4x over twelve months. This came from custom low-precision kernels (including MXFP8 attention and MLP), new 5D parallelism strategies, and topology-aware networking co-designed together across thousands of GPUs. Developers optimizing GPU training throughput at scale can follow infrastructure deep dives like this on daily.dev."}}]} ```