Meta unveils 2026 generation of its GEM ads AI: 18.9% conversion lift

One model, all surfaces

Meta announced the 2026 generation of GEM, the ads recommendation system it introduced last year. The new version moves away from separate optimizations for surfaces like Facebook, Instagram and Reels toward a single central foundation model that learns from data and signals across all of Meta's ads platforms.

Measured results

According to Meta, GEM delivered an 18.9% conversion improvement and a 12% ad click-through rate (CTR) improvement across its ad surfaces and objectives. Over the past year, the company reports additional conversion uplifts of 5.8% on Facebook, 4.1% on Instagram and 1.6% for off-site conversions among targeted users.

Scale and architecture

Meta says it built an LLM-style architecture called Kunlun, co-designed with GPU hardware. Model complexity grew 2.5x compared with earlier GEM versions, and the company observed an LLM-like scaling law at its largest pre-training scale.

The efficiency side

Because serving a model at this scale in real time is expensive, Meta used a post-training technique called LoopFM to transfer GEM's internal representations to downstream models. The company says this doubles intelligence transfer at no additional compute cost or latency. It also introduced elastic pre-computation, called Solaris.

Why it matters

Ad performance directly translates into revenue across Meta's family of apps, used by 3.6 billion people daily. By showing that LLM-style scaling in recommender systems feeds through to business results, Meta is setting a new bar for the field.