Meta Adaptive Ranking Model: Rewriting the Inference Scaling Curve to Bring LLM-Scale Models to Ads

Meta Adaptive Ranking Model directly tackles the inference trilemma, a core challenge in ad recommendation. As models grow larger and more complex, compute and memory costs surge, yet a service delivered to billions of people worldwide must still maintain sub-second latency and high cost efficiency.

To solve this, instead of applying the same model to every request, Meta introduced intelligent request routing that selects the optimal model based on the request's context and intent. As a result, the design was changed to provide more sophisticated understanding while maintaining the response speed and operational efficiency of the ad service.

The core innovations are three-fold.

• Inference-efficient model scaling: By restructuring around requests, it handles LLM-level complexity on the order of O(10 GFLOPs) while maintaining bounded latency on the order of O(100 ms).
• Model/system co-design: Architecture and execution optimizations tailored to hardware characteristics push MFU up to 35% across diverse device environments.
• Reimagined serving infrastructure: A multi-card GPU architecture and hardware-specific optimizations make it possible to serve models at a scale of O(1T) parameters.

In terms of inference efficiency, instead of processing each user-ad pair individually as before, dense user signals computed only once per request are reused across multiple candidate ads. This approach reduces redundant computation and memory bandwidth burden through Request-Oriented Optimization, Request-Oriented Computation Sharing, and In-Kernel Broadcast, and also processes long user behavior sequences on a per-request basis for reuse.

The architecture has evolved into Wukong Turbo. Building on the existing Wukong architecture, it adds a No-Bias approach, delegation that distributes small parameters, offloading from FSDP to DDP, and sparsity-based simplification to reduce the instability and network overhead that occur in deep models.

Feature preprocessing bottlenecks were also cleaned up. Preprocessing that used to run on client CPUs was moved to remote GPU hosts, and by using a tuple-based format and GPU-native kernels, Top-K complexity was reduced from O(N log N) to O(N). Adding data compression and client flow restructuring on top of this eliminated the latency penalty.

On the hardware side, selective FP8 quantization and graph/kernel specialization were applied. FP8 is used only on layers that are not sensitive to minor precision loss, shared-input operations are fused, and Grouped General Matrix Multiply and horizontal fusion bundle small operations into dense kernels to boost execution efficiency.

Large-scale embedding operations were also redesigned. Given the nature of ad recommendation, which deals with sparse categorical features, embedding tables can grow to terabyte scale. This system adjusts hash size according to feature sparsity, prunes unused embeddings, and reduces memory usage through unified embeddings, where multiple features share a single table. Larger-scale embeddings are distributed via multi-card sharding to go beyond the physical limits of a single GPU.

Operational stability was also strengthened. Multi-stream downloading and remote caching load models in under 10 minutes, and auto-scaling based on streaming multiprocessor utilization adjusts resources to match traffic fluctuations. Meta stated that after launching this model on Instagram in Q4 2025, it recorded ad conversions +3% and ad click-through rate +5% among targeted users.

Looking ahead, the focus shifts from individual optimizations toward more autonomous infrastructure. Longer behavior sequences, lower precision, agentic optimization frameworks, and incremental in-place weight updates for near-instantaneous model freshness were presented as the next steps.