More articles 8 min read
Meta Andromeda 2.0: what Meta changed in October 2026
Andromeda 2.0: what Meta published on October 28, 2026, what changed between 2024 and 2026, the official dates, and what it means for your creatives.
Contents
- What is Meta Andromeda 2.0?
- What changed between Andromeda 2024 and 2026?
- How does Andromeda 2.0 work?
- When did Andromeda 2.0 roll out?
- Andromeda, GEM, Generative Recommender: who does what?
- Is interest targeting useless now?
- What Meta did not say about Andromeda 2.0
- What to do in your accounts with Andromeda 2.0
On October 28, 2026, Meta published a long technical post on its engineering blog about the new generation of Andromeda, its ad shortlisting system. On Reddit and LinkedIn, people already call it “Andromeda 2.0”, and the advice that comes with it goes further than what Meta wrote.
This page only uses Meta sources: the October 28, 2026 post, the December 2, 2024 post, the prepared remarks of its quarterly results and its Meta for Business pages, read on November 9, 2026. For the basics, the difference between a concept and an iteration or how many creatives to launch, see our Meta Andromeda guide.
What is Meta Andromeda 2.0?
Andromeda 2.0 is the 2026 generation of Andromeda, Meta’s ads retrieval system. Retrieval is the first stage of delivery: for every possible impression, it narrows tens of millions of eligible ads down to a few thousand candidates. Ranking models then sort those candidates, and the auction picks the ad that is shown.
Meta Andromeda · delivery
Three stages before the screen
Andromeda is the first: it narrows tens of millions of eligible ads down to a few thousand candidates.
- 1 Retrieval Andromeda Tens of millions of eligible ads a few thousand
- 2 Ranking GEM and the ranking models A few thousand candidates the most relevant
- 3 Auction Ad value The top-ranked ads one ad shown
The name “Andromeda 2.0” does not come from Meta. Its post talks about “the latest iteration” of Andromeda and “the 2026 Meta Andromeda”. It describes a deep change to the model and the servers, not a new Ads Manager feature: there is nothing to turn on.
What changed between Andromeda 2024 and 2026?
Meta sums up the change in one line: the 2026 generation brings near ranking-level performance to retrieval scale. The figures published in the two engineering posts:
| Andromeda 2024 | Andromeda 2026 | |
|---|---|---|
| What gets evaluated | A hierarchical index that focuses compute on the most relevant nodes | Every eligible ad, across Facebook and Instagram |
| Throughput | Tens of millions of ads per second per GPU card | Over one billion per second per GPU server, 100 times more |
| User-side model | A single model | An ensemble of specialized models, the “experts” |
| Model power | 10,000 times more capacity | Near ranking-level accuracy, for about 300 times fewer operations per inference than the 2024 model |
| Advertiser targeting | Checked on hosts separate from GPU scoring | Checked on the same GPU server, with about 100 times less network traffic |
| Published result | +6% recall, +8% ad quality on selected segments | +2.4% conversions on Facebook since 2025 |
Andromeda 2024 → 2026
Every eligible ad gets a score
What changes: the 2024 index focused compute on the most relevant nodes, the 2026 generation scores every eligible ad.
2024
A hierarchical index: compute focuses on the most relevant nodes.
2026
Every eligible ad across Facebook and Instagram gets a score.
- > 1 billion ads per second per GPU server
- 100× the throughput of the 2024 system
- ≈ 300× fewer compute operations per inference
Why did Meta need that much speed? Its post gives the reason: generative AI creative tools have helped advertisers create “hundreds of millions” of new ads. The 2024 system evaluated tens of millions of ads per second per GPU card; the demand became several billion.
For advertisers, Meta promises “better customer matches and more conversions”, without changing anything in their campaigns, and within the same latency budget as before.
The “300” you see in some summaries is not a 300-millisecond latency: it is how many times less compute the 2026 model uses per inference than the 2024 one.
How does Andromeda 2.0 work?
The post describes three changes in the model, and one in the servers.
Experts on the user side. The 2024 model read the person with a single model. The 2026 one uses several specialized models, which Meta calls experts. They read four families of signals: the person’s sequence of interactions with ads, their user cohort, LLM-derived semantic signals, and their engagement across Meta’s surfaces.
Real-time context on the ad side. Each candidate ad is read together with the serving context of the moment.
A gate for each ad. A gating network routes each candidate ad to the experts that are most relevant to it. That is what lets Meta add experts, and therefore precision, without a proportional increase in compute.
Andromeda 2026 · the model
Experts read the person
A gating network routes each candidate ad to the experts that matter for it.
What Meta reads about the person
- Their sequence of interactions with ads
- Their user cohort
- LLM-derived semantic signals
- Their engagement across surfaces
Experts
- A
- B
- C
- D
Each ad only activates the useful experts: more precision, without multiplying compute.
Candidate ads
- Grüns Talking head, 9:16 video
- Hydratis In-store demo
- Lymphoria Product static
Read with the real-time serving context.
Everything on the same server. Targeting checks, scoring and top-candidate selection now run on the same GPU server. Meta says these checks apply “complex advertiser-defined targeting logic” across tens of feature categories, and update in real time as ads are created, modified or expire.
That last point answers a common question: the targeting you set still applies. It decides which ads are eligible for a person; Andromeda then chooses among them.
When did Andromeda 2.0 roll out?
Meta has not given a single rollout date. October 28, 2026 is the date of the engineering post. Meta measures its gains there “since 2025”, and its Q3 2026 results, published a week earlier, already reported Andromeda improvements: more model complexity and throughput, for 1.4% more conversions on Facebook.
Andromeda has kept evolving since launch. The milestones Meta has published:
Meta Andromeda · timeline
What Meta published, step by step
- H2 2024 Meta launches Andromeda with NVIDIA Retrieval models 10,000 times more complex, and 8% higher ad quality on the objectives tested. Meta, Q4 2024 results
- December 2, 2024 The first engineering post +6% recall and +8% ad quality on selected segments, across Facebook and Instagram. Meta Engineering
- Q2 2025 Facebook Reels come into scope More relevant candidates and nearly 4% higher conversions on Facebook mobile Feed and Reels. Meta, Q2 2025 results
- Q3 2025 Retrieval and early ranking combined One model for retrieval and early-stage ranking: +14% ad quality on Facebook surfaces. Meta, Q3 2025 results
- Q4 2025 Three chip families Andromeda runs on Nvidia, AMD and MTIA, Meta’s own chip. Its compute efficiency nearly triples. Meta, Q4 2025 results
- Q2 2026 A first generative model in retrieval Meta introduces Meta Generative Recommender and deploys a first generative model into its ads retrieval system. Meta, Q2 2026 results
- Q3 2026 +1.4% conversions on Facebook Andromeda scales model complexity and throughput. Generative retrieval reaches Facebook Feed: nearly 2% more conversions across targeted segments. Meta, Q3 2026 results
- October 28, 2026 The 2026 generation, a.k.a. “Andromeda 2.0” Every eligible ad gets a score, at over one billion ads per second: 100 times the 2024 throughput. +2.4% conversions on Facebook since 2025. Meta Engineering
Andromeda, GEM, Generative Recommender: who does what?
Andromeda is only one part of Meta’s ads system. The other names that come up in its 2025 and 2026 publications:
| System | Its role, according to Meta | What Meta has published |
|---|---|---|
| Andromeda | Retrieval: a few thousand candidates out of tens of millions of ads | +1.4% conversions on Facebook in Q3 2026 |
| Meta Generative Recommender | LLMs reason across ad content and the person’s preferences to predict the best ad | Nearly +2% conversions across targeted Facebook Feed segments, in Q3 2026 |
| GEM | The ads ranking model | +1.7% conversions for offsite conversion campaigns and +9.6% for app ads, in Q3 2026 |
| Lattice | The architecture that predicts ad performance by consolidating many small, separate models into larger ones (introduced in May 2023) | With GEM, a more than 6% increase in conversion rate for landing page view ads, in Q1 2026 |
One point of vocabulary matters here. The October 28 post says nothing about analyzing the images, video or audio of creatives. The system Meta describes as reasoning across ad content is Meta Generative Recommender.
Is interest targeting useless now?
The October 28 post says nothing about interest targeting. It only confirms that advertiser targeting keeps being applied before scoring.
Meta’s position on the topic is older: in April 2025, it wrote that creative diversification, rather than niche targeting, had become the best lever to find the most relevant audiences. Andromeda 2.0 does not change that message. Narrow targeting shrinks the pool of people in which the system can find those who will respond to each ad.
What Meta did not say about Andromeda 2.0
| What you read | What Meta actually published |
|---|---|
| “Andromeda 2.0 is the official name” | Meta talks about “the latest iteration” and “the 2026 Meta Andromeda” |
| “Andromeda 2.0 analyzes the image and audio of every creative” | The October 28 post does not say so; Meta Generative Recommender is the system that reasons across ad content |
| “The update happened on October 28” | That is the date of the post; the published gains run since 2025 |
| “You need one ad set per persona” | No account structure rule in Meta’s Andromeda publications |
| “You need X creatives per week” | No number tied to Andromeda, in 2024 or in 2026 |
| “Andromeda 2.0 responds in 300 ms” | The “300” means 300 times less compute per inference, not a latency |
| “The attribution window changes” | The post says nothing about attribution |
What to do in your accounts with Andromeda 2.0
Andromeda 2.0 compares more ads, more finely, for each person. It does not add diversity to your creatives: it can only choose among what you give it. Meta’s guidance on this dates from December 2025 and still holds: ads that look or feel alike are treated as variations of the same creative, and what the system learns is shared at the creative level. Meta recommends creatives that differ in look, feel, storyline and message.
In practice, five habits follow:
- Diversify concepts, not files. Ten versions of the same video are still one creative for Meta. How to build real concepts, and iterate on them, is in our Meta Andromeda guide.
- Name each creative by concept and iteration. It is the only way to read results concept by concept. A Meta Ads naming convention exists for exactly that.
- Group the formats of one creative into a single ad. The 9:16 and 4:5 versions of a video make one ad, not two: see multi-format creatives.
- Launch new concepts together, on a steady rhythm. The Meta Ads bulk upload guide explains how to load a whole batch in one pass.
- Decide on automatic enhancements yourself. Advantage+ creative options change your creatives at delivery: here is how to set them.
Apogee’s bulk upload reads the concept and iteration in each file name, groups the formats of one creative into a single ad, and launches the batch in the testing structure you pick, on Meta, TikTok and Snapchat.
Andromeda 2.0 does not ask anything new of you. It compares more ads, more finely: what makes the difference is still what you give it to compare.
Meta sources
- Meta Engineering, the 2026 generation of Andromeda (October 28, 2026)
- Meta Engineering, the launch of Andromeda (December 2, 2024)
- Meta, Q3 2026 results prepared remarks (October 21, 2026)
- Meta, Q1 2026 results prepared remarks (April 29, 2026)
- Meta for Business, Demystifying creative diversification (December 16, 2025)
- Meta AI, Lattice (May 11, 2023)