What is Meta Andromeda?

Meta's engineering team describes retrieval as the first stage of a multi-stage ads recommendation system. Retrieval narrows tens of millions of ad candidates to a few thousand relevant ones, and it has to do this quickly because it processes about a thousand times more ads than the stages after it. Larger ranking models then choose the final ads. Meta's help centre describes that final choice as an auction on total value.

Andromeda is the retrieval stage. An ad that does not make the shortlist for a person is never scored by the ranking models for that impression, so it cannot win that auction.

The article on how the Meta ad auction works explains the auction factors. Andromeda changes the step before them, which is the choice of ads considered at all.

Meta's December 2024 post calls Andromeda its proprietary machine learning system design for retrieval in ad recommendation. Meta says its earlier retrieval systems applied only limited personalisation because they relied on isolated model stages and many rule-based heuristics. Andromeda replaces that with a custom deep neural network and a hierarchical index that organises ads in layers. Meta designed the models, the serving software and the hardware together, including its own Meta Training and Inference Accelerator chips.

Meta reported these results from deploying Andromeda across Facebook and Instagram:

• A 6% improvement in recall, which measures how many of the relevant ads retrieval keeps.

• An 8% improvement in ads quality on selected segments.

• A 10,000 times increase in model capacity for personalisation.

• Ten times better inference efficiency and more than three times more queries handled per second from end to end.

Meta repeated the ads quality figure in an April 2025 post for advertisers, which says Andromeda has led to an 8% improvement in ads quality. Meta also said it planned to move the model toward an autoregressive loss function, which it expects to produce a more diverse set of ad candidates.

The December 2024 post frames Andromeda as support for Advantage+, which it says automates budget allocation, audience targeting and bid adjustments. The post notes that over one million advertisers used Meta's generative AI tools to create more than 15 million ads in a single month, and that Meta expects the number of ad creatives in its system to keep growing. Meta says Andromeda's hierarchical index is how retrieval handles that volume.

Meta's April 2025 post for advertisers makes the link to creative explicit. It says that with AI-enabled tools, the focus has shifted from niche targeting to creative diversification as the best lever for finding relevant audiences. More creative volume and variety give the system more eligible ads to personalise with, and Meta says Andromeda is what lets it process that extra information.

Meta's creative advice started before the Andromeda post. In November 2024 Meta recommended keeping at least 20 diversified ads in an Advantage+ shopping campaign and putting 20% to 30% of the overall ad budget toward testing new ads. Advantage+ shopping campaigns are now called Advantage+ sales campaigns.

Meta has not published rules for advertising under Andromeda. Its product changes and guidance since 2024 favour broad audiences and more varied creative. Ads Manager now starts sales, app promotion and leads campaigns with Advantage+ audience, placements and campaign budget switched on. Meta's best practices for Advantage+ sales campaigns ask for a wide variety of diverse creative assets.

Meta's help centre also warns about the opposite problem. Its page on ad volume says that when an advertiser runs too many ads at once, each ad delivers less often, fewer ads exit the learning phase and more budget is spent before delivery is optimised. The same page suggests fewer ads per ad set with more creative assets inside each ad, since one ad can hold up to ten.

One reading of the two pieces of guidance together is that Meta wants clear differences between the ads it can choose from, without delivery spread thin across many near-identical ads. That reading is ours, not Meta's.

Much published advice on Andromeda comes from advertisers and agencies rather than Meta, so it is interpretation, not documentation. Jon Loomer, a long-running Meta ads educator, summarised the common view in an August 2025 article:

• Make ads differ in concept, angle, format or persona, not in small copy or visual tweaks, so the system has distinct options to match to different people.

• Keep the structure simple, with one campaign per goal and few ad sets, and do most of the work at the ad level.

• Keep ad sets broad, because the ad itself now does much of the targeting.

• Expect more than six ads per ad set to work. Loomer says Meta dropped its old guidance of six ads per ad set in 2025 and links the change to Andromeda, while noting that Meta never confirmed that link.

These readings fit Meta's statements, but Meta has not published how much weight retrieval gives to creative features. Claims that Andromeda reads each image to choose an audience, or that a fixed share of each new ad must look different, are practitioner theories. Meta has said that GEM, a ranking model, learns from ad content including a representation of the creative.

Andromeda is one of several large models Meta has described for ad delivery. The others work on ranking, the stage after retrieval.

• Meta Lattice, announced in May 2023, is a model architecture that predicts ad performance across many datasets and goals that hundreds of separately trained models used to cover. Meta says it shares learning across surfaces such as Feed, Stories and Reels and across goals such as clicks and conversions.

• GEM, the Generative Ads Recommendation Model, is what Meta calls the largest foundation model for recommendation systems in the industry. Meta described it in November 2025. GEM does not serve ads itself. It passes what it learns to other ads models, and Meta credits it with a 5% increase in ad conversions on Instagram and 3% on Facebook Feed in the second quarter of 2025.

• The Adaptive Ranking Model, described in March 2026, routes each ad request to a ranking model of suitable size so Meta can serve much larger models within its time limit. Meta reports a 3% increase in ad conversions and a 5% increase in click-through rate for targeted users since its Instagram launch in the fourth quarter of 2025.

Meta has not published one diagram that connects these systems. Its GEM post describes passing knowledge to downstream models but does not name Andromeda, so any claim that GEM directly trains Andromeda goes beyond what Meta has said.