Meta says the latest version of Andromeda, its ads retrieval system, scores every eligible ad at more than one billion ads per second on a single GPU server.
The claim comes from an Engineering at Meta post published October 28. Retrieval is the first stage of Meta’s ads pipeline. It narrows a large pool of eligible ads to a short list, and a later ranking stage scores the survivors. The post says the new design brings model quality close to ranking level at retrieval scale, using about 300 times fewer inference-time compute operations than the 2024 Andromeda model, a 100x throughput lift on the 2024 system.
Meta gives a reason: generative AI creative tools have helped advertisers create hundreds of millions of new ads, which grew the candidate pool the system has to score. The post says running index matching, scoring and top-K selection on one GPU cut network traffic by about 100x.
Why it matters
The advertiser claim in the post is that performance improves automatically, because the system matches ads across a larger candidate pool within the same latency budget. Meta reports that Andromeda has delivered a 2.4% increase in conversions on Facebook since 2025. Within two weeks Meta has now published conversion figures for three parts of its ad stack: this one, the 18.9% figure in Meta Reports GEM Ads Model Lifted Conversions 18.9% Across Its Ad Surfaces and the 0.35% figure in Meta Reports a 0.35% Conversion Lift From GenRecs Tokenization.
One original insight
In our reading, a media planner should not add 2.4%, 18.9% and 0.35% together. Each is Meta’s own measurement of one component, and the posts do not say how the components overlap. What the Andromeda post does change is the cost of creative volume. If the system can evaluate hundreds of millions of ads at this speed, a large creative library costs Meta little to score, which leaves the advertiser’s test plan as the limit on how many variants get a fair trial.
Source: Engineering at Meta
