GEM Training: How Meta Doubled the Efficiency of Its LLM-Scale Ads Foundation Model

**Meta**’s **Generative Ads Recommendation Model (GEM)** is the foundation model behind ads recommendations on **Instagram** and **Facebook**. Meta Engineering reports that GEM now trains at **LLM scale** on several thousand latest-generation GPUs, with end-to-end training efficiency doubled to **20–25% Model FLOPs Utilization (MFU)** while training FLOPs scaled **4x**.

The operational claim is joint: more compute and better use of that compute. **MFU** measures how much of theoretical model FLOPs are realized in training. Reaching **20–25% MFU** at multi-thousand-GPU, LLM-scale ads training means the stack is converting a larger share of peak FLOPs into useful work rather than leaving capacity idle on communication, pipeline bubbles, or underutilized kernels.

For engineers building large recommendation or foundation-model systems, the numbers matter more than the product label. Ads ranking at this scale is not a small-model fine-tune: it is multi-thousand-GPU training with the same efficiency pressure as LLM pretraining. Doubling E2E efficiency while increasing training FLOPs **4x** is a concrete capacity lever—more model scale or more experiment throughput without a proportional jump in cluster hours.

In market terms, this sits inside Meta’s core ads stack on **Instagram** and **Facebook**, where recommendation quality is revenue-critical. An LLM-scale ads foundation model trained more efficiently is a competitive training-system story as much as a modeling one: whoever runs larger recommendation models at higher MFU can iterate faster or train richer models on the same hardware budget.

Practical takeaway: treat **MFU** and E2E training efficiency as first-class metrics when scaling recsys foundation models, not only peak GPU count. Watch whether Meta publishes more on the systems work behind the **2x** efficiency gain and the **4x** FLOPs scale-up—those details are what other teams can reuse when moving ads or ranking models into the multi-thousand-GPU regime.