Meta details engineering behind doubling training efficiency (MFU) of its GEM ads recommendation foundation model while scaling training compute 4x over 12 months

65Useful signal

Meta AI Research published a technical blog post describing how it improved end-to-end training efficiency of GEM (Generative Ads Recommendation Model), Meta's ads recommendation foundation model, from roughly 10-12% to 20-25% Model FLOPs Utilization (MFU), while scaling total training FLOPs 4x over the past 12 months. This was achieved via custom recommendation kernels (Jagged Flash Attention, GDPA, BlockAttention), mixed ultra-low-precision training (MXFP8), and topology-aware 5D parallelism combined with network/memory co-design.

infrastructureeconomicscapability

highAug 3, 2026

Was this useful?

What Happened

Meta AI Research published a technical blog post detailing engineering improvements to GEM, its ads recommendation foundation model. The company reports raising end-to-end training efficiency (Model FLOPs Utilization) from roughly 10-12% to 20-25%, while scaling total training compute 4x over the past 12 months. The gains came from custom kernels (Jagged Flash Attention, GDPA, BlockAttention), ultra-low-precision training (MXFP8), and topology-aware 5D parallelism with network and memory co-design.

Why It Matters

This is internal infrastructure powering ads ranking on Instagram and Facebook, not a product or public release, so the direct impact is Meta's own cost efficiency and ability to scale GEM further. For outside engineers and researchers, it offers directional lessons on adapting LLM-style infrastructure (parallelism, low-precision training) to sparse, embedding-heavy recommendation workloads, which is a genuinely hard and under-documented problem. Competitors and infra teams at other large recommendation-system operators (e.g. other ad platforms, e-commerce, search) are the most likely to act on this.

What Is Noise

All efficiency figures are self-reported by Meta with no external benchmark, code release, or third-party reproduction, so the 2x MFU claim should be treated as a company-approved number, not an independently verified one. No code, weights, or detailed ablations are released, so "how Meta doubled efficiency" is more a high-level engineering narrative than something others can replicate directly.

Watch Next

Whether Meta or others publish a peer-reviewed paper or open-source reference implementation of Jagged Flash Attention, GDPA, or BlockAttention
Any measurable change in Meta's ads infrastructure costs or capital expenditure guidance in upcoming earnings calls that analysts tie to training efficiency gains
Whether competing platforms (Google, Amazon, TikTok/ByteDance, Pinterest) publish similar recommendation-model infrastructure disclosures in the next 6-12 months, suggesting an industry-wide shift
Any follow-up Meta disclosure quantifying downstream effects, such as ad relevance, revenue per user, or GEM model size increases enabled by the efficiency gains

Score Breakdown

Positive Scores

Evidence Quality

17/20

Concreteness

13/15

Real-World Impact

11/20

Falsifiability

6/10

Novelty

7/10

Actionability

5/10

Longevity

7/10

Power Shift

2/5

Noise Penalties

Vagueness

-1

Speculation

-0

Packaging

-2

Recycling

-0

Engagement Bait

-0

Reasoning: First-party Meta AI engineering post with specific, measurable claims (10-12% → 20-25% MFU, 4x training FLOPs in 12 months) and named techniques (Jagged Flash Attention, GDPA, MXFP8, topology-aware 5D parallelism), making it strong primary evidence and highly concrete. Real-world impact is genuine but bounded: this is proprietary internal infrastructure powering Meta's ads stack, with no released code or product, so outsiders gain directional design lessons rather than something they can deploy. Falsifiability is limited since MFU figures are self-reported with no external reproduction path, and there is mild self-promotional packaging, though the technical density keeps hype low.