TL;DR
- Engineering at Meta said on 28 October 2026 that Andromeda now scores every eligible ad and surfaces top candidates at over 1 billion ads per second on one GPU server, a 100x lift on the 2024 system.
- Meta says Andromeda has delivered a 2.4% increase in Facebook conversions since 2025, but gives no baseline, method, time window or Instagram figure.
- Meta says the new retrieval model uses about 300x fewer inference FLOPs than the 2024 model, and moving targeting, scoring and selection onto one GPU cut network traffic about 100x.
- The post announces no new Ads Manager setting or report, and contains no data for Thailand or Southeast Asia.
Meta said on 28 October 2026 that the new version of Andromeda, its ads retrieval system, now scores every eligible ad with close to ranking-level model quality and surfaces the top candidates at more than one billion ads per second on a single GPU server, which Meta calls a 100x lift over the 2024 Andromeda system. In the same Engineering at Meta post, Meta states that "Since 2025, Meta Andromeda has delivered a 2.4% increase in conversions on Facebook", but it gives no baseline, measurement method or time window for that number.
The article is a technical engineering post with 12 bylined authors, published in the ML Applications section of Engineering at Meta. It announces no new setting in Ads Manager and no new report. Its value for advertisers is that it explains a stage of ad delivery that is normally invisible: the step that decides which ads are even considered before the auction ranks them.
What Andromeda does in Meta's ad delivery
Meta describes Andromeda as its machine learning system for personalized ads retrieval, the first stage of the ads recommendation pipeline. Retrieval 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 candidates in fine detail.
Put simply, an ad has to make it through retrieval before ranking can judge it. Meta says the deep modelling used at the ranking stage has historically been too expensive to run at retrieval scale, so retrieval relied on lighter models. The 2026 work is about bringing retrieval quality close to ranking quality without a matching rise in compute cost.
Targeting still applies. Meta's post says the retrieval pipeline includes "index matching", which it defines as enforcing advertisers' targeting rules. Meta adds that index matching evaluates advertiser-defined targeting logic across tens of feature categories and has to handle real-time updates as ads are created, modified or expire during the day. Index matching runs before scoring in Meta's pipeline, so "every eligible ad in the candidate pool", as Meta puts it, refers to ads that pass that matching step for a given person, not every ad running on Meta.
Why Meta rebuilt ads retrieval
Meta gives two reasons. The first is the size of the candidate pool. The 2024 deployment reimplemented retrieval and grew model capacity by four orders of magnitude, Meta says, and since then "generative AI creative tools have helped advertisers rapidly create hundreds of millions of new ads", growing the ads candidate pool "to an unprecedented scale". The post does not give the current size of that pool.
The second is hardware. The 2024 system evaluated tens of millions of ads per second per GPU card. Meta says new business demand required multi-billion ads per second per card, and at that rate the limit is no longer arithmetic. It is memory bandwidth inside the GPU and network traffic between machines. Meta says intermediate data between GPU steps alone added up to terabytes per second of memory traffic.
The three changes, explained for advertisers
Meta describes three linked changes to the model and the system that runs it. None of them is something an advertiser configures, but they explain what Meta means by "near ranking-level" retrieval.
A retrieval mixture-of-experts model
Meta replaced a single "user tower" with an ensemble of specialised user towers that act as experts. Meta says these draw on sequential ad engagement events, user-cohort representations, LLM-derived semantic features and cross-surface co-engagement signals. On the ad side, a joint encoding includes real-time serving context. An attention-based gating network then routes each candidate ad to the most relevant user experts. Meta says the result approaches ranking-level model performance at about 300x fewer inference-time FLOPs than the 2024 retrieval model.
One fused GPU kernel per user and ad
Meta fused the model steps and the system steps for each user-ad pair into a single GPU kernel. Intermediate results stay in GPU registers and shared memory instead of being written out to high-bandwidth memory and read back. Meta says this fused kernel runs at over 10 billion ads per second on a single NVIDIA GH200.
Targeting, scoring and selection on the same GPU server
Previously, index matching and top-K selection ran on separate CPU hosts that sent candidates over the network to remote GPUs for scoring. Meta moved index matching, scoring and top-K selection onto the same GPU. The request is now a compact payload of user features and a targeting query, and only the final top-K results come back. Meta says this reduced network traffic by about 100x. Moving index matching to the GPU meant redesigning the index for parallel execution while keeping production targeting rules intact.
The numbers in Meta's post
Every figure in the table comes from Meta's 28 October 2026 post. None has been independently verified, and most are engineering throughput measures rather than advertiser outcomes.
| Metric | Meta's stated figure |
|---|---|
| Retrieval throughput, 2026 production system | Over 1 billion ads per second on a single GPU server, scoring every eligible ad |
| Change versus the 2024 Andromeda system | 100x lift |
| 2024 system throughput | Tens of millions of ads per second per GPU card |
| Inference compute versus the 2024 retrieval model | About 300x fewer FLOPs |
| Fused kernel throughput | Over 10 billion ads per second on one NVIDIA GH200 |
| Network traffic after moving the pipeline onto one GPU | Reduced by about 100x |
| Gating kernel after the NVIDIA work | 5.3x faster: 2.4 billion to over 10 billion ads per second |
| Facebook conversions | +2.4% since 2025; no baseline, method or window stated |
The 2.4% conversion claim: what is and is not stated
Meta's sentence is short: "Since 2025, Meta Andromeda has delivered a 2.4% increase in conversions on Facebook." The post does not say what the 2.4% is measured against. It does not say whether the number comes from a holdout test, an A/B comparison or a model estimate, whether it covers one Andromeda update or several, or whether it is total conversions, conversions per impression or something else. It names Facebook only and gives no Instagram figure.
It also differs from a number Meta gave earlier this month. Relevant Audience's 23 October 2026 report on Meta's Q3 2026 results recorded Meta's statement that improvements to Andromeda lifted conversions on Facebook by 1.4%. The two figures may cover different periods or different changes. Neither source explains its method, so they should not be added together or compared as if they measured the same thing.
Even taken at face value, a platform-wide figure says nothing about a single account. It is an aggregate across many advertisers, and it should not be used to forecast any one advertiser's conversion rate or to explain a change in one account's results.
What changes for advertisers, and what does not
There is nothing to switch on. Meta describes the benefit as "better performance automatically" from matching ads across a larger candidate pool within the same latency budget. The post has no new control, no opt-in and no new report. Meta also says plainly what it gains: increased business from advertisers and lower costs, as each GPU evaluates more ads per second.
Analysis, framed as checks rather than promises:
- Targeting still decides eligibility through index matching. A narrow audience limits which people an ad can be retrieved for. The post does not tell advertisers to broaden or narrow targeting.
- Meta links the growth in candidates to generative AI creative tools and Advantage+ automation. A bigger pool means more ads competing at the retrieval stage, your own included. The post does not claim that adding more creative variants improves any advertiser's results.
- Measure with your own data. If an account shifts after this change, compare against your own conversion records and controlled tests, not against Meta's 2.4%.
- Do not pin a one-week dip or lift on Andromeda. The post says the system is "in production" but gives no date for when the 2026 version went live.
Meta says there is headroom left
Meta and NVIDIA began a joint project in early 2026 to push each part of the retrieval pipeline toward hardware limits. NVIDIA contributed optimised GPU kernels, and Meta integrated and benchmarked the full pipeline. Each component gained 2x to 5x over its baseline, according to Meta. The gating kernel reached about 85% of the GPU's peak memory bandwidth. INT8 quantisation gave up to a further 2x over 16-bit kernels with accuracy on par, and index matching passed 100 billion ads per second. Meta is now extending the work to NVIDIA Grace Blackwell and says the pipeline "is not yet at its ceiling".
Analysis: this suggests retrieval will keep changing under live campaigns without any announcement inside Ads Manager. Engineering posts such as this one are where those changes get described, which is why they are worth reading even when they contain no new setting.
What Meta's post does not say
- No date for when the 2026 version went into production.
- No current size of the ads candidate pool.
- No figure for Instagram, Threads or WhatsApp.
- No country or regional results, and no mention of Thailand or Southeast Asia.
- No baseline, method or time window for the 2.4% figure.
How this fits Relevant Audience's earlier Meta coverage
Relevant Audience has not run a dedicated post on Andromeda before. On 15 October 2026 it covered Meta's 13 October engineering post on GEM, Meta's Generative Ads Model, and the missing context around its 18.9% conversion claim. On 23 October its Q3 2026 results report recorded the 1.4% Andromeda figure from Meta's earnings. The 28 October post is new in three ways: it explains the retrieval stage itself, it publishes the throughput and compute figures above, and it introduces the 2.4% since-2025 figure.
What this means for Thai marketers
Meta's post contains no Thailand data, so nothing in it is a Thailand-specific result. The change applies to how ads on Facebook and Instagram are shortlisted for each person, and Meta does not describe any regional rollout. For Thai advertisers, the practical points are the ones above: keep targeting rules deliberate, read Meta's percentages as Meta's own unverified estimates, and judge results on your own conversion data.
If you want help checking your conversion setup or structuring campaigns, Relevant Audience runs Facebook and Instagram ads management and broader paid social campaigns for brands in Thailand.
Frequently asked questions
What is Meta Andromeda?
Andromeda is Meta's machine learning system for personalized ads retrieval, the first stage of its ads recommendation pipeline. It narrows the pool of eligible ads to a small set of candidates in milliseconds before Meta's ranking stage scores them in detail.
Did Andromeda really increase conversions by 2.4%?
Meta says so but does not show how it measured it. The 28 October 2026 post states a 2.4% increase in conversions on Facebook since 2025 with no baseline, test design, time window or Instagram figure, so the claim cannot be checked or applied to a single account.
Do advertisers need to change any settings?
No. Meta announced no new setting, opt-in or report, and says advertisers get better performance automatically. Targeting rules still apply through index matching.
Does "scores every eligible ad" mean targeting no longer matters?
No. Meta's pipeline runs index matching, which Meta describes as enforcing advertisers' targeting rules, before ads are scored. Which ads are eligible for a person still depends on how they are targeted.
Is there any data for Thailand?
No. The post gives no country or regional results and does not mention Thailand or Southeast Asia.
The short version: Meta has made its first filtering step much faster and closer in quality to its ranking models, and it says advertisers benefit automatically. The engineering figures are detailed. The advertiser figure is a single sentence, and it is the one to treat with the most care.
