Meta has updated the technologies that determine who sees ads on Facebook and Instagram. In October 2026, the company introduced a new Andromeda architecture, a scaled-up GEM model, and the GenRecs Tokenization system. Together, they help select ads more accurately, account for user behavior, and understand the content of new creatives even before performance data accumulates. However, these changes do not guarantee more impressions for every ad or automatically require advertisers to restructure their campaigns.
Key Takeaways
- On October 28, 2026, Meta introduced a new Andromeda architecture capable of more than 1 billion ad evaluations per second on a single GPU server.
- The candidate retrieval system’s throughput increased by approximately 100 times compared with Andromeda in 2024.
- GEM combines learning from signals across Meta’s advertising surfaces and uses longer sequences of user behavior.
- GenRecs Tokenization represents ads using semantic tokens, helping assess new products and ads without a history of interactions.
- Meta reports conversion gains in its own measurements, but those results do not predict performance for every ad account.
- The updates do not automatically change the rules for campaign structure, creative testing, or budget management.
Why “Andromeda 2.0” is an inaccurate name
In industry discussions, Meta’s three technical publications from October are often grouped under the name “Andromeda 2.0.” However, the company has not introduced a single product under that name.
In fact, these are separate components of the advertising infrastructure, which Meta described in individual announcements:
- October 13: a new generation of the Generative Ads Model (GEM).
- October 19: GenRecs Tokenization — a semantic representation of ads, products, and other entities.
- October 28: the next generation of Andromeda, an initial ad candidate retrieval system.
These technologies interact, but they perform different tasks. Andromeda narrows an enormous pool of ads down to the most promising candidates. GEM helps advance prediction models, while GenRecs provides a better understanding of advertising content.
Andromeda now evaluates every eligible ad
Andromeda operates at the retrieval stage — the initial selection of ads. Before an ad is shown, Meta has to determine which ads from a large pool are potentially relevant to a particular person. Only then do the selected candidates move on to more complex ranking models.
In its 2026 version, Meta says the system can evaluate all ads eligible for a particular request instead of using a less precise approach that considers only some candidates.
To achieve this, the company rebuilt the architecture around GPUs and adopted a Mixture-of-Experts approach. Instead of one general-purpose model, it uses specialized components that account for different aspects of a user’s interests and behavior.
The new system combines targeting eligibility checks, ad evaluation, and selection of the best candidates in a single GPU pipeline. This reduces data transfers between processors and servers.
In a production configuration, Meta reports more than 1 billion ad evaluations per second on a single GPU server. That is roughly 100 times the throughput of the 2024 system. A separate optimized computing component can process more than 10 billion ads per second on a single GPU, but that is a different metric from the throughput of the complete pipeline.
Will all creatives receive more impressions?
No. This is a common but incorrect interpretation of the update.
Andromeda can now evaluate a broader set of eligible ads with high accuracy. However, after evaluation, the system still selects only the most relevant candidates.
As a result, a creative that previously failed to reach the next stage because of initial retrieval limitations may have a better chance of being evaluated. But this does not mean guaranteed impressions or an even distribution of budget across ads.
Meta also reports that Andromeda has driven a 2.4% increase in Facebook conversions since 2025. This is a result from the company’s internal measurement, not a promise of similar gains for every advertiser.
GEM: more context about user behavior
The Generative Ads Model is Meta’s large-scale foundation model for recommendations. Its role is to learn from signals across the advertising ecosystem and transfer that knowledge to models used when serving ads.
In 2026, GEM is moving from separate models for different advertising surfaces to a central architecture trained on data from Facebook, Instagram, Reels, and other environments.
The new Kunlun architecture can work with longer sequences of user actions and more complex relationships between behavior and advertising offers.
However, the claim that GEM now analyzes exactly twice as much user history needs clarification. Meta reported a doubling in the efficiency of transferring knowledge to other models, not a universal doubling of user history for every ad request.
According to Meta’s internal data, advances in GEM delivered an 18.9% improvement in conversions and a 12% increase in CTR in company measurements across different advertising surfaces and objectives. These figures cannot be directly applied to the results of an individual campaign.
GenRecs helps the system understand new ads
One challenge in ad recommendations is dealing with new products and creatives that have not yet accumulated enough impressions, clicks, or conversions.
A traditional system may use unique object identifiers and gradually learn from interaction histories. GenRecs Tokenization adds a semantic representation: models analyze text, images, and other signals and convert them into compact sets of tokens.
For example, a new ad for hiking shoes might share semantic characteristics with previous ads for similar products. The system draws on patterns it has already learned, even if the specific new product has no interaction history of its own.
This does not mean that Meta automatically identifies the exact audience for any creative or that a new ad no longer needs performance data.
According to Meta, GenRecs Tokenization, combined with GEM and the Adaptive Ranking Model, produced a 0.35% increase in conversions. The measurements were taken in February 2026, so this is not a separate result from the October 19 launch.
What changes for media buyers?
Meta’s October publications primarily describe infrastructure changes. The company did not announce new mandatory Ads Manager settings, a universal number of creatives per ad set, or new rules for scaling.
Nevertheless, these technical changes support several practical conclusions.
1. Creative diversity should be meaningful
If Meta understands ad content better, it makes sense to test not only colors, backgrounds, or minor copy changes, but also different offers, product use cases, and messages for potential customers.
However, simply increasing the number of ads does not guarantee better results. Additional creatives should provide new information about what resonates with the audience.
2. New creatives may receive a relevant evaluation sooner
GenRecs reduces the reliance of initial predictions on the history of a specific ad. This could be useful for new products, niche offers, and campaigns with large catalogs.
But a faster relevance assessment does not mean a campaign will exit the learning phase sooner. These are different processes, and Meta has not announced the removal of the learning phase.
3. There is no need to restart stable campaigns
The Andromeda update does not require manual activation. There is no evidence that duplicating campaigns or creating new ad sets improves the performance of the new models.
If a campaign delivers an acceptable CPA or ROAS, changing it solely because of Meta’s technical announcement is not advisable. Restarting it can make performance comparisons harder and affect optimization.
4. Do not switch off creatives more frequently without sufficient data
The claim that weak ads will necessarily start spending their budgets faster after the update is unsubstantiated.
Decisions to turn off ads should be based on campaign results: spend, conversion volume, CPA, ROAS, and lead quality. A few impressions or a handful of expensive clicks do not always provide enough information.
5. Broad targeting should be tested, not treated as mandatory
The development of Andromeda, GEM, and Advantage+ strengthens the role of automated audience selection. But that does not mean all manual targeting settings have become irrelevant.
Geography, legal restrictions, audience exclusions, and specific business requirements remain important for certain products.
UAGEEK discussed the changing balance between automation and manual control in its article on Andromeda and Advantage+ in media buying.
How to assess the update’s impact on your campaigns
Meta has not provided a tool that allows advertisers to see exactly how Andromeda evaluated a particular ad. The impact of infrastructure changes can therefore be assessed only indirectly, through advertising results.
- Record your baseline metrics. Compare CPA, ROAS, CPM, CTR, frequency, and conversion volume across comparable periods.
- Control for external factors. Account for seasonality and changes in competition, budget, offer, and landing page.
- Test different ad messages. Compare creatives with substantively different messages, not just cosmetic variations.
- Do not change everything at once. If you update creatives, targeting, and budget at the same time, identifying the cause of changes in results will be difficult.
- Cross-check conversions. For affiliate campaigns, compare Meta’s data with your tracker and the affiliate program’s postback data, taking attribution differences into account.
It is important not to attribute every drop or increase in performance to Andromeda. Without a controlled experiment, that would be speculation.
Meta is also developing audience settings at the ad creative level. UAGEEK has already covered Creative Audience Pairing — choosing an audience for each creative. This is a separate Ads Manager feature, not part of October’s Andromeda announcement.
What Meta’s numbers actually mean
The metrics in Meta’s technical publications need to be read in the proper context.
- 1 billion ads per second: the throughput of Andromeda’s production GPU pipeline, not the number of ad impressions served to users.
- 100-fold speedup: a comparison of the retrieval system’s throughput with the 2024 architecture.
- 2.4% increase in conversions: the result Meta attributes to Andromeda’s development on Facebook since 2025.
- 18.9% increase in conversions: Meta’s internal metric for GEM improvements across different advertising surfaces and objectives.
- 0.35% increase in conversions: an internal result for GenRecs Tokenization in combination with other models, measured in February 2026.
These figures were obtained under different conditions and should not be added together. Nor do they guarantee corresponding growth in conversions or profit in a particular ad account.
Conclusion
The October updates show the direction in which Meta is developing its advertising system: more accurate initial retrieval, deeper use of behavioral signals, and a better understanding of new ad content.
For media buyers, this is not a new set of Ads Manager rules, but a change in the technologies beneath its interface. The practical takeaway is to work with meaningfully different creatives, maintain high-quality conversion signals, and evaluate results based on campaign economics.
Neither automatically increasing the number of creatives, nor switching ads off more frequently, nor restarting campaigns has a proven advantage solely because a new version of Andromeda was released.
Sources for verification
Sources for verification
Engineering at Meta, October 28, 2026. Official technical overview of the new Andromeda architecture, Mixture-of-Experts, collaboration with NVIDIA, throughput of more than 1 billion ads per second, and results for advertisers.
The next generation of Meta Andromeda
Engineering at Meta, October 13, 2026. Primary source on the development of GEM, the Kunlun architecture, long behavioral sequences, knowledge transfer between models, and performance metrics.
How Meta scales the Generative Ads Model
Engineering at Meta, October 19, 2026. Description of GenRecs Tokenization, semantic representations of new ads, the knowledge transfer mechanism, and measured conversion gains.
How GenRecs understands ad content
Engineering at Meta, December 2, 2024. The original technical announcement of Andromeda, allowing a comparison of the 2024 and 2026 architectures.
The original Meta Andromeda architecture