# What is Meta Lattice?

> Meta Lattice is one AI model that predicts how an ad will perform across many goals and surfaces, replacing many smaller ones. What Meta said and what it means.
- **Author**: Marxx
- **Published**: 2026-11-22
- **Category**: Meta
- **URL**: https://marxx.ai/library-glossary/meta-lattice

---

Meta Lattice is one AI model that predicts how an ad will perform across many goals and surfaces. It covers work that many smaller models did before, each on one job. Meta announced it on 11 May 2023 ([Meta AI blog](https://ai.meta.com/blog/ai-ads-performance-efficiency-meta-lattice/), checked 2026-10-20).

## The problem it fixes

Meta's ads system had grown to hundreds of models. Meta trained each one separately, for its own goal, such as ad quality or conversion rate. That was slow to improve and costly to run. A new surface with little data also struggled to learn. Meta calls this the cold start problem.

## How it works

Lattice learns from all of those goals and surfaces together and shares what it learns between them. Meta says it also handles delayed signals. A person may click now and buy days later, and the model can use both.

Meta reports that the model has trillions of parameters and trained on hundreds of billions of examples from thousands of data domains.

## A simple picture

Say Meta once ran one small model for each pair of surface and goal. With 3 surfaces (Feed, Stories, Reels) and 3 goals (clicks, video views, purchases), that is 9 models, each learning alone. With one shared model, a lesson from Reels purchases can help Stories purchases too. This is a simplified picture, not Meta's real count.

| | Many small models | One shared model |
|---|---|---|
| Models to train | 9, one per surface and goal | 1 |
| What each one learns from | Its own surface and goal | All 9 together |
| A new surface with little data | Starts almost from zero | Borrows from the others |

## What Meta says it gained

In early use on Instagram, Meta says sharing knowledge across surfaces and objectives improved results for advertisers. It reported about an 8% improvement in ads quality. That is Meta's own figure, from 2023.

## What it means for you

You cannot change Lattice. It runs inside Meta's system. Meta says it designed Lattice for a world with less detailed data, so the signals you do send matter. Send clean purchase events through the [Conversions API](/glossary/conversions-api) and let ad sets get enough results to leave the [learning phase](/glossary/learning-phase).

Meta's newer engineering posts focus on [GEM](/glossary/meta-gem-model). Another named system, [Andromeda](/glossary/meta-andromeda), narrows the pool of ad candidates before ranking.

## FAQ

### Is Lattice the same as Andromeda?

No. [Andromeda](/glossary/meta-andromeda) is the retrieval stage that narrows the pool of ad candidates. Lattice predicts how an ad will perform.

### Do I need to do anything about Lattice?

No. Meta's post describes no advertiser setting for it. Run enough varied creative and send clean conversion data. That is good practice whichever model Meta uses.

### Where does this come from?

From Meta's own post of 11 May 2023, linked above. The 8% figure and the model sizes are Meta's claims, not independent tests.

---
- [More Meta articles](https://marxx.ai/library-glossary/category/meta)
- [All articles](https://marxx.ai/library-glossary)