Meta trains one GEM model on every ad surface and passes its gains to user-facing models
Meta says its central GEM ads model now learns from every ad surface and has delivered an 18.9% conversion gain. The post says a model that size is too costly to serve live, so the gains pass to other models through post-training.
The Engineer · Build desk
Ranking of the additional conversion uplift GEM contributed for targeted users over the past year, per Meta: Facebook 5.8%, Instagram 4.1%, off-site conversion 1.6%.
Added conversion uplift from GEM, targeted users, past year
| Rank | Item | Value | Claim |
|---|---|---|---|
| 1 | 5.8% | 15 | |
| 2 | 4.1% | 15 | |
| 3 | Off-site conversion | 1.6% | 15 |
What happened
- Before this generation, each Meta ad surface and optimization objective ran its own independent prediction system, across apps that 3.6 billion people use daily.
- The new GEM jointly predicts clicks and conversions across all surfaces and across advertiser objectives from off-site conversions to mobile app installs.
- A post-training technique named LoopFM moves GEM's internal representations and sequential historical memory into downstream models.
- Meta says computation sharing in LoopFM doubles the transfer gained from the previous version's knowledge distillation, at no additional compute cost or latency.
- Meta reports 2.5x model complexity growth from earlier GEM versions, with an LLM-like scaling law observed at its largest pre-training scale.
Why it matters
- constraint Because live serving at GEM's scale is described as prohibitively expensive, the reported gains measure GEM plus the transfer step into user-facing models, not GEM alone.
- cost The compute bill sits in training. The post's own section title puts it at 4x total FLOPs across thousands of GPUs. A team that wants to copy the design would have to pay that bill first.
- exposure One upstream model feeding every surface is one upstream dependency. We'd expect a bad training run to reach every surface that inherits from it.
Meta's post says its post-training techniques provide a larger latency budget without requiring real-time evaluation of all ad candidates [2]. The cost problem is stated in the same terms: running a model this large live costs too much to be practical, and it adds latency [3]. In our view the design follows from that. The large model cannot sit in the request path, so its representations do [3][4]. Meta adds an elastic pre-computation piece, Solaris, to harvest the upstream gains [6]. The excerpt does not describe how Solaris is scheduled.
The central model is where learning happens. The post says the 2026 generation moves off isolated per-domain optimization for Facebook, Instagram and Reels [1]. It also says GEM's output goes to downstream models [4], without saying whether those models are still built per surface.
Meta says plain transformer scaling did not carry over to recommenders, because of massive heterogeneous features and shifting distributions [11]. Its answer is Kunlun, an LLM-style architecture co-designed with the latest GPU hardware and built for long-context learning over user engagement history [10].
Both headline gains, the conversion figure and a 12% click-through improvement, are Meta's own, measured against a baseline the excerpt leaves undefined, and span Meta's surfaces and objectives [14]. The per-platform uplifts are 5.8% on Facebook, 4.1% on Instagram and 1.6% on off-site conversion, over the past year and for targeted users [15].
Meta also says representations learned on data-rich click tasks transfer into data-sparse conversion tasks [9]. For these numbers to transfer to another stack, that stack would need the same click-rich, conversion-sparse structure. It would also need Meta's claim of no added compute cost or latency from LoopFM to hold on its own serving hardware [5].
What to watch
- Whether Meta defines the baseline and time window behind the 18.9% conversion and 12% CTR figures.
- How Meta measured the no-added-cost claim for LoopFM, and how Solaris pre-computation is scheduled.
- The post's training-stack section, for how the 4x FLOPs and doubled end-to-end efficiency were measured.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence35
- Adoption70
- Hype gap+25
- Incentives70
- Confidence40
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
The 2026 generation of GEM shifts from isolated per-domain optimizations, such as Facebook, Instagram and Reels, to a single central foundation model that learns from data and signals across all of Meta's ads platforms.
- [2]
Post-training techniques provide a larger latency budget without requiring real-time evaluation of all ad candidates.
- [3]
Meta says serving a model at this scale in real time is prohibitively expensive and drives up latency, compromising user experience and advertiser ROI.
- [4]
LoopFM, a post-training technique, transfers GEM's internal representations and sequential historical memory to downstream models.
- [5]
Through computation sharing, LoopFM doubles the intelligence transfer together with the knowledge distillation deployed in the previous version, at no additional compute cost or latency.
- [6]
Meta introduced elastic pre-computation, called Solaris, to harvest the gains from the upstream model.
- [7]
Each Meta ad surface and optimization objective previously had its own independent prediction system, across apps used by 3.6 billion people each day.
- [8]
GEM jointly predicts clicks and conversions across all surfaces, including Facebook and Instagram, and across advertiser objectives from off-site conversions to mobile app installs.
- [9]
Generalizable representations learned on data-rich click tasks transfer into data-sparse conversion tasks.
- [10]
Meta built Kunlun, an LLM-style architecture co-designed with the latest GPU hardware, to make larger, deeper models more cost-efficient and to enable long-context learning over user engagement history.
- [11]
Recommender system models do not exhibit transformer-like scaling properties because of massive heterogeneous features and shifting distributions.
- [12]
The fourth section of the post is titled 'Scaling Total Training FLOPs by 4x While Doubling E2E Training Efficiency Across Thousands of GPUs'.
- [13]
Meta observed an LLM-like scaling law at its largest pre-training scale, with 2.5x model complexity growth from earlier versions of GEM.
- [14]
GEM has delivered 18.9% conversion improvement and 12% ad click-through rate improvement across Meta's ad surfaces and objectives.
- [15]
GEM additionally contributed 5.8%, 4.1% and 1.6% conversion uplift on Facebook, Instagram and off-site conversion for targeted users over the past year.
Sources
1 independent publisher whose own reporting we read for this story.
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Ads recommendation systemsFollow
- Foundation models beyond languageFollow
- ML training infrastructureFollow