Meta Unveils Adaptive Ranking Model to Scale AI-Powered Ads with Unprecedented Efficiency

Meta has developed the Adaptive Ranking Model, a groundbreaking AI system designed to serve large language model (LLM)-scale recommendation models for its Ads platform while maintaining sub-second latency and cost efficiency. The innovation addresses the inference trilemma—balancing model complexity, computational demands, and real-time performance at global scale—by dynamically routing requests to the most effective models based on user context and intent.

Overcoming the Inference Trilemma

Scaling Meta’s Ads Recommender models to LLM complexity posed a critical challenge: balancing increased computational needs with the platform’s strict latency and cost constraints. The Adaptive Ranking Model resolves this by replacing a rigid, one-size-fits-all inference approach with intelligent request routing, ensuring each user query is processed by the optimal model without compromising speed or efficiency.

Since its deployment on Instagram in late 2025, the system has delivered measurable improvements, including a 3% increase in ad conversions and a 5% rise in click-through rates for targeted users. These gains underscore its ability to enhance ad relevance while maintaining computational efficiency.

Core Innovations Behind the Breakthrough

The Adaptive Ranking Model achieves LLM-scale performance through three key advancements:

• Inference-Efficient Model Scaling: By adopting a request-centric architecture, the system reduces redundancy in LLM-scale computations, enabling sub-second latency even with complex models. This shift transforms scaling costs from linear to sub-linear, optimizing hardware utilization.

• Model/System Co-Design: The model is hardware-aware, aligning its architecture with underlying silicon capabilities to boost model FLOPs utilization (MFU) to 35% across diverse hardware environments.

• Reimagined Serving Infrastructure: A multi-card GPU infrastructure breaks single-device memory limits, allowing O(1T) parameter scaling—unprecedented for real-time recommendation systems.

Technical Pillars of Efficiency

The system’s efficiency stems from three technical pillars:

• Request-Oriented Optimization: Computes high-density user signals once per request instead of per ad candidate, eliminating redundancy.

• Structural Throughput Maximization: Architectural refinements stabilize deep models and minimize network bottlenecks.

• Latency Optimization: Offloads feature preprocessing to GPUs and streamlines execution paths to neutralize complexity overhead.

Together, these innovations ensure personalized ad experiences while maximizing advertiser value and computational efficiency at Meta’s global scale.

Read more: engineering.fb.com