Andromeda 2.0 - Meta did it once again

Hellooo,

It’s Andromeda again. And I tried to bring the news as soon as I could.

On October 28, Meta's engineering team published the first full technical update to Andromeda since the system launched in December 2024. Andromeda is the retrieval layer, the step that decides which ads are even eligible to compete, before ranking or the auction runs. This post is Meta explaining what they rebuilt and why.

What actually changed between the two versions

The 2024 system increased retrieval model capacity by roughly 10,000x and cut feature-extraction latency about 100x versus the CPU-based setup it replaced. 

The 2026 rebuild goes further on throughput: the production pipeline now scores every eligible ad and surfaces top candidates at over 1 billion ads per second on a single GPU server. Meta states this as a 100x lift over the 2024 system, using a retrieval model with roughly 300x fewer inference FLOPs.

The architectural change matters more than the throughput number. Instead of one general model of user interest, Meta introduced a Mixture-of-Experts setup, separate specialized models scoring sequential ad engagement, shifting interests over time, semantic/contextual signals, cross-surface activity, and cohort similarity. 

A gating network picks which experts apply to each ad-user pair. Meta and NVIDIA also moved indexing, scoring, and candidate selection onto one unified GPU pipeline, which is what made the throughput jump possible without blowing the latency budget.

What this establishes, and what it doesn't

It establishes that Meta is evaluating a larger, more varied ad pool earlier in the funnel, using more granular signal types, without slowing delivery. That's a real engineering shift, there's no new Ads Manager setting attached to it.

On results, Meta states Andromeda "has delivered a 2.4% increase in conversions on Facebook" since 2025, with no baseline, measurement window, or method disclosed. 

Separately, Meta's Q3 2026 earnings commentary attributed a 1.4% Facebook conversion lift to Andromeda improvements over roughly the same period.

Where the data layer still does the work

A bigger, faster retrieval pool raises the ceiling on what Meta can consider for your ad. It doesn't change what Meta is told to optimize toward. That instruction still comes from the conversion events you feed in. 

If the only signal reaching Meta is "purchase happened" or "lead submitted," a faster retrieval system will keep finding more of whatever is cheapest to label purchase or lead, not necessarily the one with the margin or close rate that matters to the business.

This is the practical read for Advantage+ and broad-targeting setups generally. Broader retrieval is the default behavior of the system, not a setting you control. 

The lever you do control is the specificity and accuracy of the event data defining the outcome, deduplicated identity across browser and server, which purchase was high-AOV versus low-AOV, which lead was sales-qualified versus a form fill. 

That's a data pipeline question, independent of whatever Meta changes on the retrieval side.

What to check before assuming this changes your account

• Confirm event deduplication is actually working, matching event_id across browser and server-side events, not just firing both.

• Check which conversion event each ad set is optimizing for, and whether that event reflects margin or just occurrence.

• Confirm whether offline/CRM conversions are being sent back at all, or whether Meta is only ever seeing the online half of the funnel.

• If performance moves after October 28, compare against your own conversion records and a controlled test, not against Meta's 2.4% (or 1.4%) figure, since neither comes with a stated baseline.

The question this analysis can answer is whether your retrieval pool got better, it can't tell you whether your optimization events do. That second question is the one worth auditing this week.

Meta’s official release

Book a 1PD Ops demo

Start a 14-day free trial