22% ROAS improvement. 23.7x higher retrieval throughput. 291-523x faster filtering than CPU-based systems.
On May 26, 2026, Meta deployed SilverTorch to its global ads fleet — and if you are still managing Meta campaigns the same way you were a week ago, you are already optimizing against the wrong model.
What SilverTorch Is
Most algorithm changes tweak weights or add signals. SilverTorch is different — it is a complete architectural replacement of the ads retrieval stack.
The old system (Andromeda microservices): Separate services for indexing, filtering, scoring, and reranking — running mostly on CPUs. Multiple handoffs, version drift between models.
SilverTorch: A single unified PyTorch model running entirely on GPUs (NVIDIA Grace Hopper + Meta MTIA chips). The item index and eligibility rules are now tensors and layers within the model itself — not separate services. This is the "Index as Model" paradigm.
Three technical breakthroughs:
- GPU Bloom Index — 291-523x faster candidate filtering than CPU inverted indexes
- Fused Int8 ANN Kernel — custom nearest-neighbor GPU kernel, up to 14.7x faster than standard implementations
- OverArch and Value Model — neural reranking layers that optimize for multi-task scoring (click vs. conversion vs. downstream value) at retrieval time
What This Means for Ad Delivery
Root cause: SilverTorch shifts Meta ads from demographic matching to semantic matching. Your ad creative is now the primary targeting signal — not your audience settings.
Key findings:
- Creative-first targeting: The GPU-native stack processes high-dimensional creative signals in real-time. Who sees your ad is increasingly determined by what your creative signals — not which interest buckets you checked.
- Real-time index freshness: New ads can be retrieved and matched within minutes of upload. No more 24-48 hour index lag for new creative.
- 5.6-6% improvement in retrieval recall on engagement tasks — confirmed in Meta's internal evaluations.
- CPM volatility is expected during the transition window. As the auction recalibrates around semantic signals, CPMs will fluctuate. Do not panic-pause campaigns.
Performance Impact
+22% ROAS for Advantage+ campaigns — reported in Meta's early evaluations. The engine was purpose-built to amplify Advantage+ automation. Manual targeting campaigns are now competing against a system that was not designed with their manual signals in mind.
Recommended Response
1. Lead with creative diversity. Provide a wide variety of creative angles. More diverse creative = broader semantic surface area = more efficient audience discovery.
2. Test Advantage+ now. SilverTorch was built to run Advantage+ natively. This is no longer optional as a consideration — it is mandatory due diligence.
3. Feed the Value Model. Quality conversion data via Conversions API is now more valuable than ever. The OverArch layer does multi-task optimization — give it downstream value signals (purchase value, LTV, repeat purchase).
4. Audit interest-targeting redundancy. Overly granular interest filters can now restrict the semantic reach SilverTorch is trying to expand.
"The pipeline was not broken; it was state-unaware." SilverTorch fixes that permanently by collapsing the pipeline into the model. Every creative signal you upload now goes directly into the retrieval engine.
Big difference.
Founder CEO/CTO
