Meta today published the technical design of the second generation of Andromeda, the system that decides which ads on Facebook and Instagram reach the ranking stage, and claimed it has delivered a 2.4% increase in conversions on Facebook since 2025. According to the company, the rebuilt retrieval engine now scores every eligible ad in its candidate pool at more than one billion ads per second on a single GPU server - a figure Meta describes as a 100x lift over the version it launched in December 2024.

In Short

Before Meta picks which ad a person sees, a first system quickly shortlists candidates from an enormous pile of ads, and Meta has rebuilt that system so it can check every eligible ad, more than one billion of them every second on one server. That matters to anyone advertising on Facebook or Instagram, because an ad that is not shortlisted never reaches the auction, and Meta says the change has lifted Facebook conversions by 2.4% since 2025. Advertisers get no new setting or report from this, and the 2.4% figure comes from Meta alone, with no baseline or method attached.

What Meta published

The post appeared today on Engineering at Meta, the company's technical blog, filed under ML Applications and Production Engineering. Its byline lists 12 engineers: Haiyu Lu, Shuo Gu, Yang Yang (Tony), Jingwen Ren, Tianchen Ji, Qing Zhang, Wentian Guo, Haocheng Wu, Siqi Zeng, Yunchen Pu, Liang Wang and Chonglin Sun. An acknowledgements section names 122 further contributors and thanks the wider team behind the system.

This is new material, not a resurfaced version of the original. Meta first described Andromeda on December 2, 2024, in an engineering post that presented it as the retrieval engine behind Advantage+ automation. Today's post, titled "Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design," lists that 2024 article as related reading and describes what has changed since. The company calls the current system "the 2026 Meta Andromeda."

The timing fits a pattern. Fifteen days ago, Meta published a long account of GEM, its central ads ranking model, in which it claimed an 18.9% conversion improvement across its ad surfaces. Nine days ago came a post on GenRecs tokenization for ads recommendation. With today's publication, Meta has now described both ends of its ad delivery pipeline in October: GEM on the ranking side, Andromeda at retrieval.

What retrieval does, and why it was the constraint

Meta describes Andromeda as "Meta's machine learning system for personalized ads retrieval, the first stage of the ads recommendation pipeline." Retrieval, according to the post, "narrows a vast pool of eligible ads down to a small set of strong candidates in milliseconds; a later, more computationally expensive ranking stage then scores those survivors in fine detail." When the system launched, Meta put the scale of that narrowing at tens of millions of ad candidates reduced to thousands per request.

The two-step design exists because of cost. "Historically, the deep modeling quality of the ranking stage has been too costly to apply at retrieval's scale," the post states. The claim Meta makes today is that it has closed most of that gap: in production, according to the company, the system "scores every eligible ad with near ranking-level model performance at retrieval scale."

Why rebuild a system that launched less than two years ago? Meta points to its own creative tools. "Since then, generative AI creative tools have helped advertisers rapidly create hundreds of millions of new ads, growing Meta's ads candidate pool to an unprecedented scale," the post states. Meta does not give a current size for the pool. Its earnings disclosures show the adoption behind it: more than 8 million advertisers were using at least one of Meta's AI creative tools in the first quarter of 2026, up from 4 million at the end of 2024, and the company said nearly 10 million advertisers used its generative AI creative tools in the third quarter.

The 2024 build, according to today's post, ran from "custom AI hardware with hierarchical indexing and GPU-native operators to a purpose-built ML model, growing model capacity by four orders of magnitude." Four orders of magnitude is 10,000x, the capacity figure Meta gave at launch.

Three problems at billion scale

Meta frames the redesign around three constraints. The first is model quality. "Ads retrieval has an outsized ROI on incremental quality: Small improvements in retrieval model performance can drive large gains in advertiser value," the post states. To keep improving at "a sustainable compute cost," Meta says, it had to move beyond the 2024 design to an architecture that raises performance "without a proportional increase in costs."

The second is throughput. "The 2024 Meta Andromeda system evaluates at tens of millions of ads per second per GPU card. Meeting the new business demand requires a throughput at multi-billion ads evaluated per second per card in the system." That per-card figure is new; the December 2024 post reported a more than 3x gain in end-to-end queries per second but gave no count of ads evaluated per second. At billions of ads per second, arithmetic stops being the limit. Intermediate tensors passed between GPU kernels "account for terabytes per second of memory traffic," according to Meta, and shipping ads between CPU hosts and remote accelerators "saturates the network link."

The third is hardware headroom. Meta wanted to know "How close are we to the hardware ceiling, and how much room remains for future model growth?" That question, according to the post, "motivated a joint collaboration with NVIDIA."

A mixture-of-experts model for retrieval

The model change borrows from large language models. The latest Andromeda "introduces three synergistic architectural innovations that bring large language model (LLM)-inspired mixture-of-experts (MoE) into the retrieval," according to Meta. In a mixture-of-experts design, a model contains many specialised sub-networks, and a routing component activates only a few of them for each input. Capacity can grow without every request paying for the whole model.

Meta's headline claim for the architecture is about the cost of inference, the compute spent each time the model scores an ad. According to the post, "model performance approaches ranking-level in a typical retrieval-then-ranking system, at approximately 300x fewer inference-time FLOPs than the 2024 Meta Andromeda retrieval model." FLOPs, floating-point operations, are the basic unit of compute a model consumes.

The three changes sit in different parts of the model.

User towers. Meta "transitioned from a single user tower to an ensemble of specialized user towers as experts." These draw on "sequential ad engagement events, user-cohort representations, LLM-derived semantic features, and cross-surface co-engagement signals," according to the post. Meta does not say which surfaces feed the cross-surface signals or which language models produce the semantic features.

Candidate ads tower. Here Meta uses "a joint encoding strategy that incorporates real-time serving context," so that the interaction model receives ad representations that reflect the circumstances of each request.

Interaction layer. The component that matches users to ads "evolved from dense neural networks into an MoE attention-based gating network that dynamically routes candidates across the most relevant user experts." The caption of Meta's architecture diagram describes an "ad-aware MoE router, activating only the top-K most relevant experts to produce the user-ad relevance score."

Meta puts inference throughput for the model at "over 10 billion ads per second on a single GH200 GPU," the Nvidia Grace Hopper Superchip on which the 2024 system also launched. The company argues the efficiency leaves room for "richer feature interactions, additional expert sub-networks, and deeper architectures, without a proportional increase in GPU cost."

What does "near ranking-level" mean in measurable terms? The post does not say. It gives no recall, precision or ads quality figure for the new architecture, and no comparison against the ranking models it is said to approach. The 2024 post, by contrast, reported a 6% recall improvement and an 8% ads quality improvement on selected segments.

One kernel, one server

The system changes are as large as the model changes. According to Meta, the 2026 version rests on "a single GPU kernel fusing both model and system workloads per user-ad interaction and a strategy to co-locate per user request logics with per user-ad GPU kernels in a single GPU box."

The fused kernel

For each user-ad pair, Andromeda runs a short sequence: gathering embeddings, executing the interaction model and scoring. In the 2024 launch, according to the post, "only operations within the model execution were fused." The roughly 300x reduction in compute made it possible to fold the non-model system work into the same kernel.

The reason is memory, not maths. "At billions of ads per second, even a small intermediate tensor per ad multiplies into terabytes per second of high bandwidth memory (HBM) traffic," the post states. The fused kernel executes the full per-ad computation in one pass. "Intermediate results stay in GPU registers and thread block shared memory and are never materialized to HBM," according to Meta.

Moving targeting onto the GPU

Fusing the kernel exposed the next bottleneck. Two per-request operations, index matching ("enforcing advertisers' targeting rules," in Meta's words) and top-K selection, ran on separate CPU hosts and exchanged data with the GPU scoring service over the network. "At multi-billion ads per second throughput, this traffic alone saturates the network link," the post states.

Meta's answer was to "co-locate index matching, scoring, and top-K selection on the same GPU." Instead of shipping candidate lists between machines, the system sends "a compact request payload - user features and a targeting query - to a single GPU server," which runs the whole retrieval pipeline and returns only the final top results. According to Meta, this "reduces network traffic by ~100x."

This is the part of the rebuild that touches advertiser settings most directly. "Ads index matching evaluates complex advertiser-defined targeting logic across tens of feature categories, and must support real-time index updates as ads are created, modified, or expired throughout the day," the post states. Meta says it redesigned the index for parallel GPU execution "while preserving the full fidelity of production targeting semantics." The post presents no test of that equivalence.

Combined, the two changes produce the figure in the opening of Meta's post. The co-located pipeline "processes over one billion ads per second on a single GPU server while scoring every eligible ad in the candidate pool," according to the company, which calls it "a new throughput benchmark for ads retrieval-stage models." The caption of the system diagram specifies that this covers "every eligible ad across Facebook and Instagram."

The Nvidia collaboration

In early 2026, according to the post, Meta and Nvidia began a joint project to push each component of the retrieval pipeline toward hardware limits. "NVIDIA contributed optimized GPU kernels, and Meta integrated and benchmarked the full pipeline," the post states. Each component achieved a 2x to 5x speedup over the baseline.

The results, component by component:

  • Gating network. The optimised kernel "improved by 5.3x, from 2.4 billion to over 10 billion ads per second, reaching ~85% of the GPU's peak HBM bandwidth." Meta says the gating mechanism now runs as fast as "a simple dot-product scoring kernel" on the same hardware. "This is the co-design validation: the additional model complexity is essentially free on Tensor Core hardware," the post states.
  • Quantisation. INT8 quantisation "achieved a further up to 2x throughput increase over the optimized FP16 kernels, with accuracy on par with 16-bit precision." The post does not say whether INT8 scoring runs in production.
  • Index matching. Throughput "improved over 2x, reaching over 100 billion ads per second - an order of magnitude higher than the scoring network." That margin, according to Meta, means index matching "can be re-run across multiple retrieval pathways without becoming a bottleneck."
  • End-to-end pipeline. Data movement between components, "not computation," accounted for a significant share of latency. After data layouts were redesigned, "the full pipeline benchmarks at multiple times the current production throughput on a single GPU server, with further gains expected as additional optimizations integrate into production." The text puts no number on that multiple.

Work has moved on to Grace Blackwell. Meta says it is focusing on "platform design optimizations that enable different CPU-to-GPU compute ratios," which its results so far suggest "is an important factor for Andromeda-like workloads."

The hardware story is narrower than earlier disclosures. The 2024 post described a system built on Nvidia Grace Hopper and Meta's own MTIA accelerator. In January 2026, Meta said it had extended Andromeda to run on Nvidia, AMD and MTIA chips, a change it said nearly tripled the system's compute efficiency. Today's post describes only Nvidia hardware. It does not mention AMD or MTIA, or say whether the fused single-GPU pipeline runs on either.

What the figures leave open

All of the numbers in the post are Meta's own. None has been independently audited, and several are harder to read than the headline suggests.

The 2.4% conversion figure

The conversion claim is a single sentence: "Since 2025, Meta Andromeda has delivered a 2.4% increase in conversions on Facebook." The post gives no end date for the period, no baseline, no test design and no figure for Instagram, even though the system scores ads across both apps. It also does not say which version of Andromeda produced the gain, or how much came from the mixture-of-experts architecture versus the system changes.

Meta's investor disclosures use different numbers. On October 21, the company said improvements to Andromeda, including scaling model complexity and throughput, had increased conversions on Facebook by 1.4% in the third quarter. Whether that 1.4% sits inside today's 2.4% is not stated. A year earlier, Meta said Andromeda had delivered a 14% increase in ads quality on Facebook surfaces in the third quarter of 2025, and at launch it cited an 8% ads quality improvement on selected segments. The metric has shifted from ads quality to conversions, and Meta has not published a definition of either.

The 100x comparison

Meta's 100x figure compares the 2024 system's "tens of millions of ads per second per GPU card" with the new pipeline's "over one billion ads per second on a single GPU server." The units differ. The post's diagram shows one Grace CPU and one Hopper GPU inside the server, which suggests a one-to-one comparison, but the text never states how many GPUs a production server holds.

There is also a tenfold gap between the two headline throughput figures. The scoring kernel alone runs at more than 10 billion ads per second; the full pipeline, including index matching and top-K selection, runs at more than one billion. The 10 billion figure matches the result of the Nvidia-optimised gating kernel. The production figure is the one advertisers' ads actually pass through.

Production versus benchmark

The post mixes production results with benchmarks. The one billion figure is described as production. The Nvidia results are described as benchmarks, with gains "expected" as they reach production. The post does not say when the mixture-of-experts architecture or the single-GPU pipeline went live, or whether they reached all advertisers at once.

That matters for a system that advertisers cannot see. Andromeda has no Ads Manager setting, and its effects show up only indirectly in delivery. Practitioners often label any shift in Meta delivery an Andromeda update, though Meta has never published a rollout date for one. Today's post does not change that.

Generative retrieval

There is a further gap. In July, Meta told investors it had deployed the first generative model into its ads retrieval system, and in October it said the Meta Generative Recommender had launched as generative retrieval on Facebook Feed, driving a nearly 2% increase in conversions across targeted segments. Today's post does not mention generative retrieval or the Generative Recommender. Its only reference to parallel routes is the note that index matching can run "across multiple retrieval pathways." How Andromeda and the generative system divide the work is not explained.

Why this matters for advertisers

Retrieval is the gate. An ad that Andromeda does not shortlist never reaches ranking, the auction or a user's feed. A retrieval stage that scores every eligible ad, rather than an approximation of the pool, changes which ads get a chance in principle, and Meta's framing ties the change directly to advertiser outcomes. "For advertisers, this means better performance automatically from the retrieval system matching ads across a larger candidate pool, within the same latency budget," the post states.

"Automatically" is the operative word. There is nothing for an advertiser to switch on. The gains, if they arrive, come through campaigns that rely on Meta's automated delivery, which is where advertiser scepticism has concentrated. PPC Land's November 2025 analysis of reactions to Meta's AI automation cited one incrementality test in which Advantage+ generated only 17% of the conversions Meta's attribution reported. Today's post offers no incrementality data.

The post is also explicit about who else benefits. "For Meta, it means increased business from advertisers and lower costs, as each GPU now evaluates more ads per second," it states. Lower cost per evaluation reduces Meta's serving bill. What advertisers pay is set in the auction, and in the third quarter of 2026, Meta's average price per ad rose 14% worldwide while ad impressions grew 10%. Meta has attributed that price increase to ad performance gains and improved macro conditions, the same performance argument today's post makes about retrieval.

There is a circular element too. Meta's generative AI tools add ads to the candidate pool; the larger pool demands more retrieval capacity; Meta has rebuilt retrieval to handle it. "As the volume of eligible ads continues to grow, driven by Advantage+ automation and generative AI creative tools, Meta Andromeda's co-designed architecture scales to meet it," the post concludes. Meta's own products are, by its account, a main source of the load.

The infrastructure behind this is not cheap. Meta expects 2026 capital expenditures of $135 billion to $145 billion, and its infrastructure chief Santosh Janardhan has put 2026 AI spending at more than $100 billion, including the Prometheus cluster in New Albany, Ohio. The post does not put a cost on the Andromeda rebuild or say how much GPU capacity the system uses.

A stack rebuilt around language-model techniques

Andromeda's redesign follows a pattern visible across Meta's ads system for three years. Since introducing its Lattice architecture in 2023, Meta has cut the number of its ads ranking and recommendation models by about 100 through consolidation into larger models. GEM brought LLM-scale training to ranking. In April 2026, Meta described an Adaptive Ranking Model that routes requests to more expensive inference models when it estimates a higher probability of conversion. Today's post extends the same logic, mixture-of-experts routing and hardware-specific engineering, to the first stage of the pipeline.

The common thread is selective compute: models that decide, request by request, how much computation to spend. For Andromeda, Meta says that approach has made ranking-quality models affordable at a stage where they were previously too costly.

What comes next

Meta says the work is not finished. "These results confirm that the co-designed retrieval pipeline is not yet at its ceiling," the post states. With quantisation and index matching headroom, the company argues, the architecture "is positioned to absorb significantly more model complexity - deeper interaction networks, richer feature spaces, and additional retrieval pathways - without proportional infrastructure growth."

The Grace Blackwell work centres on CPU-to-GPU ratios, and Meta says it is exploring how the findings "extend to the next platform generation." The post gives no timeline for either, and no performance figures for Blackwell hardware.

For people using Meta's apps, the company promises "better ads that are more relevant to them, delivered within the same latency budget." For advertisers, it promises "better customer matches and more conversions." The evidence offered for the second claim, so far, is one sentence and one number.

Timeline

Summary

Who: Meta, through an Engineering at Meta post credited to 12 engineers, with Haiyu Lu, Shuo Gu and Yang Yang (Tony) listed first on the byline, with Nvidia contributing optimised GPU kernels. The changes affect every advertiser whose ads are delivered on Facebook and Instagram.

What: A technical account of the second generation of Andromeda, Meta's ads retrieval system. It replaces the 2024 model with a retrieval mixture-of-experts architecture using about 300x fewer inference-time FLOPs, fuses model and system work into one GPU kernel, and moves targeting-rule matching and top-K selection onto the same GPU server, cutting network traffic by about 100x. Meta says the production pipeline scores every eligible ad at more than one billion ads per second per GPU server, 100x the 2024 system, and that Andromeda has delivered a 2.4% increase in Facebook conversions since 2025. No baseline, period or method accompanies the conversion figure.

When: Published today, October 28, 2026. It follows the original Andromeda post of December 2, 2024, and a Meta-Nvidia collaboration that began in early 2026.

Where: Engineering at Meta, the company's technical blog. The system runs on Nvidia GH200 Grace Hopper hardware in Meta's data centers, with work under way on Grace Blackwell, and selects ads for Facebook and Instagram.

Why: Generative AI creative tools have added hundreds of millions of ads to Meta's candidate pool, and the 2024 design could not score that pool with ranking-quality models at acceptable cost. Meta says the rebuild delivers better matching for advertisers and lower serving costs for itself, as it spends between $135 billion and $145 billion on capital expenditures in 2026.