Andromeda is the retrieval engine inside Meta's advertising system. Each time a person opens Facebook or Instagram and an ad slot becomes available, Andromeda takes the tens of millions of ads that could in principle be shown and narrows them to a few thousand candidates. Heavier ranking models then score that shortlist and an auction picks the winner. It exists because automated campaigns and generative creative tools multiplied the number of ads in circulation, and Meta needed a first stage that could search a far larger pool, with far larger models, inside the milliseconds a page load allows.

Meta introduced it on December 2, 2024. The name has since become media-buyer shorthand for a wider set of campaign changes, a source of confusion in itself.

Where Andromeda sits in an ad request

Meta's ad delivery runs as a cascade. Retrieval comes first: it selects candidates from the full inventory of active ads that a given person is eligible to see. Ranking follows, where larger and slower models predict how likely that person is to click, convert or otherwise act on each candidate, and estimate the value to the advertiser. The auction then combines those predictions with bids and an ad quality assessment to choose what appears.

According to the December 2024 Meta Engineering post, written by eight engineers including Chonglin Sun, Nancy Yu and GP Musumeci, the retrieval stage narrows "tens of millions of ad candidates" to "a few thousand relevant ad candidates", handling about three orders of magnitude more ads than the later stages. An ad that is not retrieved is never ranked or seen, whatever its budget. In its third-quarter 2025 results, Meta described Andromeda as a system that combines retrieval and early-stage ranking, so the boundary between the two stages is not fixed.

Andromeda has no setting in Ads Manager. Its influence shows up only indirectly, in which ads in an ad set receive delivery and which receive almost none.

How the system works

Older retrieval systems commonly used two-tower networks, which turn a person and each ad into separate vectors and find the closest matches with an approximate nearest-neighbour search. The approach is cheap but limits how much the two sides interact.

Andromeda replaced that design with what Meta calls a highly customised deep neural network with sublinear inference cost, meaning the compute needed grows more slowly than the number of ads. Three mechanisms carry most of the load, according to Meta Engineering:

  • Hierarchical index. Ads are organised into layers of clusters. Retrieval descends through the tree and evaluates only the most promising branches, so it does not need to score every ad. The index and the retrieval model are trained jointly, which Meta says improves precision and recall over two-tower retrieval with nearest-neighbour search.
  • Model elasticity. The system applies heavier models to ad segments judged more valuable and adjusts model complexity and the number of inference steps in real time, depending on available capacity. Together with the hierarchical network, Meta credits this with a 10x gain in inference efficiency.
  • Hardware co-design. The model was built for Nvidia's Grace Hopper Superchip. Precomputed ad embeddings and features sit in the chip's local memory, and feature extraction moved from CPUs to Hopper GPUs.

Meta reported that the design allowed a 10,000x increase in retrieval model capacity, a 6% improvement in recall, an 8% improvement in ads quality on selected segments, more than 100x faster feature extraction than the prior CPU components and more than 3x end-to-end queries per second. All of those figures are Meta's own, from internal testing, with no published baseline or external audit.

The same post tied the project explicitly to Advantage+, Meta's automated campaign suite. Advantage+ automates budget allocation, audience selection and creative variation, and each of those increases the number of eligible ads that retrieval must consider. Meta wrote that "the number of ads creatives in Meta's recommendation systems is expected to grow significantly". Advantage+ tools can generate up to 150 creative variations from one set of inputs.

Origin and evolution

Andromeda is one of three named systems Meta has used to describe its ads machine-learning stack. Lattice, introduced in 2023, consolidates many smaller ranking and recommendation models into fewer large ones; by October 2025, Meta said it had cut about 100 models this way. GEM, the Generative Ads Recommendation Model, is a large foundation model for ranking, first detailed by Meta Engineering on November 10, 2025. Andromeda covers the retrieval end.

At launch, Meta set out two next steps: moving Andromeda to an autoregressive loss function, for faster inference and a more diverse candidate set, and running it on its own MTIA accelerator and future GPUs, with an estimated further 1,000x increase in model complexity. Both have since moved forward. In the fourth quarter of 2025, Meta extended Andromeda to run on Nvidia, AMD and MTIA chips, which together with model changes nearly tripled its compute efficiency. On July 29, 2026, Meta CFO Susan Li said Meta had "deployed the first generative model into our ads retrieval system and saw notable improvements in ads performance", according to the transcript. She did not name Andromeda or quantify the gain, and Meta has not said whether this is the autoregressive model flagged in 2024.

Meta later credited Andromeda with a 14% ads quality increase on Facebook surfaces in the third quarter of 2025, against the 8% on selected segments reported at launch. The scopes differ, and Meta has not published how "ads quality" is calculated.

Why it matters to advertisers

The practical consequence is that the ad itself has become a targeting input. If retrieval matches people to ads largely on what the ad depicts, near-identical creatives give the system one option, not ten. Meta's own advertiser guidance moved in that direction. In December 2025, Meta told advertisers that "Andromeda - our retrieval technology - quickly scans tens of millions of ads" and urged them to treat creative diversification as a "foundational best practice", varying formats, messages and personas rather than iterating on one concept, according to Social Media Today. Its 2025 festive season guide recommended at least three asset types and discouraged running several ad sets aimed at similar audiences.

That second point drives the structural change. Fewer, broader campaigns with many distinct creatives give retrieval a larger pool to work with and exit the learning phase faster, according to Meta. Advantage+ is now a large business: Meta put its annual revenue run rate above $75 billion in July 2026.

Limitations and disputes

The first dispute is over the name. Agencies and tool vendors routinely refer to an "Andromeda update", as if a single algorithm change were switched on at a known date. Meta has described Andromeda as infrastructure introduced in 2024 and improved continuously, and has never published a rollout date for advertisers. Jon Loomer, a Meta ads consultant, wrote in August 2025 that "Andromeda is about retrieval, the first step of ad delivery" and that "if nothing works, don't blame Andromeda", according to his site. Many changes blamed on Andromeda also involve GEM and Advantage+ defaults.

The second concerns creative similarity. Practitioners widely report that Meta groups visually or thematically similar ads under a shared identifier, often called an Entity ID, and treats them as one option. Minna Vainio of Precis wrote in July 2026 that ads sharing an Entity ID "compete against each other for the same limited audience pool", according to the agency's blog. Meta has not documented such an identifier, so the mechanism is inferred from delivery patterns.

The third is consolidation itself. Bram Van der Hallen called claims about single-campaign consolidation "absolute nonsense" in November 2025, arguing that "reach expands, but impact doesn't follow". The same coverage cited an incrementality test in which Advantage+ drove only 17% of the conversions Meta's attribution reported. Agency executives quoted by The Verge said clients would not trust outputs when Meta is "basically checking their own homework". Every Andromeda performance figure published so far comes from Meta.

Finally, smaller advertisers must produce more distinct concepts, and Meta's answer, generative creative tools, raises its own brand control concerns.

  • GEM. A ranking model, not retrieval. GEM scores the few thousand candidates that Andromeda returns and transfers its knowledge to smaller serving models.
  • Lattice. A model consolidation framework for ranking and recommendation, separate from the retrieval index.
  • Advantage+. The advertiser-facing campaign suite. Andromeda is infrastructure that Advantage+ campaigns rely on, not a product with settings.
  • Other Andromedas. The Andromeda galaxy (Messier 31) is the nearest large spiral galaxy to the Milky Way. In technology, Andromeda was also the reported code name for Google's planned merger of Android and Chrome OS; 9to5Google's Stephen Hall reported on June 4, 2017 that the project had been shelved, according to TechRadar.

Recent developments

October 2026 has brought adjacent changes rather than new Andromeda disclosures. On October 13, Meta published a GEM update reporting an 18.9% conversion improvement across its surfaces, a figure given without a baseline or method. On October 19, Meta opened per-creative audiences to Marketing API partners, letting advertisers attach an age range and gender to each creative inside an ad set that must itself stay open to all genders. Meta did not mention Andromeda, though the feature lets advertisers feed their own view of a creative's audience into the selection process. With the Adaptive Ranking Model and a generative model now inside retrieval, more of the choice of ad is made by models advertisers cannot inspect.

Timeline

  • June 4, 2017 (other sense) - Google's Android and Chrome OS "Andromeda" project reported shelved.
  • 2023 - Meta begins consolidating ads models under Lattice.
  • December 2, 2024 - Meta Engineering publishes the Andromeda post, reporting 10,000x model capacity, +6% recall and +8% ads quality on selected segments, built for Nvidia Grace Hopper.
  • August 18, 2025 - Jon Loomer publishes a widely read explainer framing Andromeda as retrieval only.
  • October 29, 2025 - Meta reports a 14% ads quality gain from Andromeda on Facebook surfaces in Q3 2025.
  • November 10, 2025 - Meta Engineering details GEM, its ranking foundation model.
  • November 16, 2025 - Meta's festive season guide recommends at least three asset types and fewer overlapping ad sets.
  • December 17, 2025 - Meta guidance names Andromeda as its retrieval technology and calls creative diversification a foundational practice.
  • January 28, 2026 - Meta says Andromeda now runs on Nvidia, AMD and MTIA chips, nearly tripling compute efficiency.
  • July 29, 2026 - Meta says it has deployed its first generative model into ads retrieval.
  • October 13, 2026 - Meta reports an 18.9% conversion improvement from GEM.
  • October 19, 2026 - Meta launches per-creative audiences for Marketing API partners.

Summary

Who: Meta built and operates Andromeda; its engineering team, led on the published work by Chonglin Sun and colleagues, introduced it, and every advertiser buying Facebook and Instagram ads is affected by it.

What: Andromeda is the retrieval stage of Meta's ad system, a hierarchical deep neural network that narrows tens of millions of eligible ads to a few thousand candidates for ranking and auction.

When: Meta unveiled it on December 2, 2024, extended it to Nvidia, AMD and MTIA hardware in late 2025, and added a generative model to retrieval by July 2026.

Where: Inside Meta's data centres, at the first stage of every ad request on Facebook and Instagram, behind Advantage+ and manual campaigns alike.

Why: Automated campaigns and generative creative tools expanded the pool of ads beyond what older retrieval could search with large models in milliseconds, and Meta needed a first stage that scales sublinearly with that pool.