On October 28, Meta's engineering team published the technical write-up on the next generation of Andromeda, the system that picks which ads get considered for every person, on every scroll, across Facebook and Instagram. Marketers are calling it Andromeda 2.0.
Most coverage stops at "Meta's AI got smarter." For anyone actually buying media, the important detail is narrower: the retrieval stage is no longer a rough pre-filter. It now scores every eligible ad at close to the quality of Meta's full ranking models. That changes what wins in the account.
01 /What Andromeda actually does
Meta's ad delivery runs in stages. When someone opens Instagram, there are millions of ads they could be eligible for. Ranking all of them with Meta's heaviest models would be far too slow, so the system works like a funnel:
- Retrieval (Andromeda) sorts through every eligible ad and returns a small shortlist within milliseconds.
- Ranking applies deeper, more expensive models to that shortlist to predict clicks, conversions and value.
- Auction combines your bid, the predicted action rates and ad quality to choose the winner.
The catch is that an ad that doesn't make the retrieval shortlist never reaches the auction, whatever its bid or however good its creative. Retrieval is the gate, and Meta has just made that gate much more precise.
02 /Why Meta had to rebuild it
The first Andromeda launched in late 2024 on NVIDIA Grace Hopper (GH200) chips. Meta says it grew retrieval model capacity by four orders of magnitude. Then generative AI happened to the ad inventory.
With Advantage+ and Meta's AI creative tools, advertisers have produced hundreds of millions of new ads. Each one is another candidate retrieval has to evaluate. The 2024 system handled tens of millions of ads per second per GPU. Meta's demand projections called for billions. Making the 2024 model bigger wouldn't get there, so the model, software and hardware were redesigned together.
03 /The headline numbers
The last number is the one advertisers should note. A 2.4% conversion lift across a platform of Facebook's size comes from infrastructure alone, with no change to anyone's campaigns. It shows that better retrieval feeds directly into results.
04 /The new model: retrieval mixture-of-experts
Meta took a technique familiar from large language models, mixture-of-experts (MoE), and applied it to retrieval. Instead of one large network doing everything, several specialised "experts" each learn a different view of the user, and a router selects which ones to use for each ad.
Three changes that matter
- Many user experts instead of one user tower. Separate experts learn from ad engagement sequences, cohort behaviour, LLM-derived semantic understanding and co-engagement across Meta's surfaces, so the system has several distinct views of what a person responds to.
- Ads are encoded with live context. Your creative is represented together with the real-time serving context, not as a fixed fingerprint. The same ad can score differently depending on the moment and the placement.
- An attention-based router replaces dense layers. For each candidate ad, only the most relevant experts run. Meta says this delivers near-ranking quality at a small fraction of the compute, and that after optimisation the gating costs about as much as a simple dot product.
Because the router picks experts per ad, a meaningfully different creative (new angle, new format, new hook) can be matched through a different "view" of the user and reach a different pocket of people. Ten variations of one concept probably all look the same to it. Creative diversity is now literally how the retrieval model is built.
05 /One GPU server, one pipeline
The model was only half the work. At billions of ads per second, Meta found that the bottleneck was moving data, not computing it. Two engineering changes fixed that:
- A single fused kernel. Fetching embeddings, the user-ad interaction and scoring now run in one GPU pass, with intermediate results kept on-chip instead of written out to memory.
- Co-location. Index matching, which enforces your targeting rules, used to run on CPU servers and send huge candidate lists across the network to GPUs. Now matching, scoring and top-K selection all run on the same GPU server. Only a small request goes in and only the final shortlist comes out, cutting network traffic by about 100×.
One detail for media buyers: moving index matching onto the GPU required a new index design. Meta says it still supports complex targeting across dozens of feature categories and real-time updates as ads are created, edited or switched off. Your targeting constraints, exclusions and new launches are respected immediately, just at much higher speed.
06 /Meta × NVIDIA: how close to the limit?
In early 2026 Meta and NVIDIA benchmarked the pipeline against the hardware's limits. NVIDIA supplied optimised kernels and Meta integrated and tested them. Every component improved by 2× to 5× over its baseline:
| Component | What it does | Result |
|---|---|---|
| Interaction network | MoE gating that scores user × ad | 2.4B → 10B+ ads/s (5.3×) |
| INT8 quantisation | Lower-precision gating math | up to 2× vs FP16, same accuracy |
| Index matching | Enforces targeting rules | 100B+ ads/s (2×+) |
| End-to-end | Full pipeline on one server | multiple × production |
The practical point is that Meta has room to keep making retrieval smarter (deeper networks, more experts, more signals) without its infrastructure costs growing at the same rate. Expect this layer to keep improving through 2027.
07 /What it means for advertisers
Meta's post covers engineering, not campaign strategy. Here's how we read it from the media-buying side, based on what we see across the eCommerce and lead-gen accounts we manage.
1. Every ad competes on its own merits
When retrieval was coarse, a weak ad could ride along inside a strong ad set. Now each ad is scored for each user at close to ranking quality. The best creative for a specific person is far more likely to be found, and weak creative is far more likely to be filtered out before the auction.
2. Creative is your targeting
The model reads ad content directly and matches it to user experts. The hook, visuals, offer and format effectively decide who sees an ad. Narrow interest targeting mostly limits the pool the system can search.
3. Signal quality feeds the experts
The user experts learn from engagement and conversion behaviour. Clean, deduplicated Conversions API events with strong match quality, and optimising for the event you actually care about, give the system better training data, which leads to better matching.
4. Volume without variety is noise
Meta built this upgrade partly because AI tools created hundreds of millions of ads. Having more ads is not the advantage. More genuinely different ads is, because that's what lets the router find new pockets of demand.
08 /The FRX Andromeda playbook
Here's what we're doing in client accounts this quarter:
Consolidate structure
Fewer campaigns and ad sets, each with enough budget to exit learning. Use Advantage+ Sales or broad ABO/CBO with Advantage+ audience, and let retrieval do the searching.
Build concepts, not variations
Aim for 6–10 truly distinct concepts per ad set: different angles (problem, social proof, offer, founder, comparison), different people on camera, different formats (UGC video, static, carousel, catalog).
Fix the signal layer first
Pixel + CAPI with deduplication, an Event Match Quality of 8 or higher where possible, value passed on purchases, and offline or CRM events for lead-gen. Judge results from your backend, not just Ads Manager.
Refresh on a cadence
Ship new concepts every 1–2 weeks at meaningful spend. Read concept-level results (hook rate, CTR, CPA by concept), not just individual ads, and cut losers fast.
Use exclusions, not interests
Keep hard constraints (geo, age where legally required, existing-customer exclusions) and drop speculative interest layers. Index matching still enforces your rules, so only keep the ones you truly need.
Andromeda 2.0 doesn't reward clever account hacks. It rewards a steady supply of distinct, well-made creative, backed by clean conversion data. The accounts that win in 2027 will be the ones with the best creative systems.
09 /FAQ
No. Meta refers to it as the next generation, or 2026 iteration, of Meta Andromeda. "Andromeda 2.0" is the shorthand marketers use.
No. It's already serving across Facebook and Instagram, and Meta's 2.4% conversion lift on Facebook required no advertiser action. You'll get more from it with consolidated structure, diverse creative and clean CAPI data.
Hard constraints like location, exclusions and compliance rules are still enforced at the index-matching step. Speculative interest stacking matters much less, because the model matches people to ads using their content and behaviour signals.
Count distinct concepts, not ad volume. For most accounts, 6–10 genuinely different concepts per ad set, refreshed every one to two weeks, works far better than dozens of near-identical versions.
Sources
Engineering at Meta, "Meta Andromeda: next-generation ads retrieval with GPU co-design," October 28, 2026. engineering.fb.com. All figures on this page are original FRX Media illustrations of the data and architecture Meta describes.