Meta detailed on October 28, 2026 the new generation of Meta Andromeda, the artificial intelligence system that performs the first screening of ads on Facebook and Instagram. According to the company, the production pipeline evaluates every eligible ad at more than 1 billion per second on a single GPU server — a leap Meta describes as roughly 100 times the performance of the Andromeda launched in 2024.
What Andromeda does
Before any auction, Andromeda narrows a huge pool of eligible ads down to a shortlist of candidates in milliseconds. Only then does a heavier ranking stage examine those survivors in detail. If an ad is discarded in this first screening, it never reaches the auction — which is why Meta treats retrieval as the bottleneck of ad delivery.
The problem grew with generative AI creative tools: advertisers began producing hundreds of millions of creative variations, expanding the candidate pool to an unprecedented scale, according to Meta itself.
The new model: mixture-of-experts
The 2026 architecture adopts mixture-of-experts, inspired by large language models: several specialized sub-networks share the work, and a router activates only the most relevant ones for each input. Meta moved from a single user tower to an ensemble of specialized user towers, combined candidate-ad encoding with real-time context, and replaced the interaction layer with a mixture-of-experts attention gating network.
According to the company, the result approaches ranking-level performance at the retrieval stage with about 300 times fewer inference FLOPs than the 2024 retrieval model. The model alone would exceed 10 billion ads per second on a single NVIDIA GH200 GPU — a different figure from the full pipeline, which Meta puts at more than 1 billion per second per server.
One GPU server, end to end
The other half of the change is systemic. Meta fused embedding collection, model execution and scoring into a single GPU kernel, and placed targeting-rule checks (index matching), scoring and top-K selection on the same server. Previously, some of these operations ran on separate CPU hosts and exchanged large lists over the network — traffic that, according to the company, falls by about 100 times under the new architecture.
The work was done in collaboration with NVIDIA starting in early 2026, with optimized kernels delivering 2x to 5x speedups per component. The gating network, for example, reportedly rose from 2.4 billion to more than 10 billion ads per second, reaching about 85% of the GPU's peak memory bandwidth.
The conversion figures — and the caveats
Meta says that since 2025, Andromeda has delivered a 2.4% increase in conversions on Facebook. That figure comes from the company itself, with no baseline, methodology or Instagram figure published in the post — so it should be read as a company statement, not as an independently audited result.
The numbers should not be mixed up either: on October 21, Meta told investors that Andromeda improvements had raised Facebook conversions by 1.4% in the third quarter, while the 8.3% more clicks and 15.7% more conversions in the second quarter refer to the broader ranking system, not to Andromeda alone. For advertisers, there is no new switch or report: the upgrade runs inside the infrastructure.
The 2026 Andromeda completes a trio of October Meta publications on AI advertising, alongside the GEM ranking model and the GenRecs system — a sign that the fight for attention in the feed is increasingly decided by systems advertisers cannot inspect impression by impression.
Image credit: ContentGrip — illustration about Meta Andromeda.
Sources: Engineering at Meta, “Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design” (October 28, 2026), PPC Land, ContentGrip and Marketing Dive.
Transparency: this article was produced with the help of artificial intelligence, based on cross-checking public posts and the internet sources listed below. We recommend consulting the original sources to confirm the details.
By Muse
