Meta presented a new generation of Andromeda on October 28, the system that selects candidate ads for display on Facebook and Instagram. The update changes the initial stage of advertising recommendations, before ads undergo final ranking. According to the company, the production infrastructure exceeds one billion ad evaluations per second on a single GPU server.
For businesses running advertising campaigns on Facebook and Instagram, the change sits within the mechanism that identifies eligible ads for a person. Meta aims to examine more options with a more expressive model while retaining the response time required to serve an ad. Advertiser-defined targeting parameters continue to be applied during selection.
Initial selection and final ranking perform different tasks
Andromeda operates in the first recommendation stage, known as retrieval. From a large pool of eligible ads, the system produces a smaller candidate set within milliseconds. The subsequent ranking stage applies more detailed and computationally expensive evaluations to that group.
The new architecture seeks to bring modeling quality approaching that of final ranking into initial selection. Meta describes a joint redesign of the model, processing software and the way graphics processors are used. The three components are developed together to support the volume of ads being analyzed.
The company connects the change to the expansion of advertising inventory. According to Meta, generative AI creative tools have contributed to hundreds of millions of new ads since Andromeda's initial launch in 2024. More variants also increase the amount of information the system must examine for each selection.
More specialized components for interests and context
The updated model uses a mixture-of-experts architecture inspired by large language models. Instead of a single component representing the user, Meta describes an ensemble of specialized components. They process interests that change over time and signals originating from different usage contexts.
The information listed by the company includes sequences of advertising interactions, user-cohort representations, semantic features derived with language models and engagement across platform surfaces. Ad content is combined with information about the context in which it would be served.
A routing mechanism then selects the specialized components relevant to the user–ad pairing. Meta says this organization approaches the modeling performance of the ranking stage while using approximately 300 times fewer inference-time computing operations than the 2024 Andromeda model. The comparison concerns technical architecture rather than the cost of an advertising campaign.
Fewer transfers between processors and memory
Another change brings processing stages together on the same GPU infrastructure. Previously, some operations for each request ran on CPU servers, with candidate lists sent across the network to accelerators evaluating the model. At very large volumes, these transfers became a constraint on the system.
In the new structure, eligibility-rule matching, scoring and selection of the best candidates are brought together on the same GPU. The system sends a compact request containing the necessary information and receives the final shortlist. Meta estimates an approximately 100-fold reduction in network traffic associated with this reorganization.
The team also combined intermediate operations within a single processing kernel. Temporary results remain in the graphics processor's registers and shared memory instead of repeatedly being written to and read from the accelerator's main memory. The technical aim is more efficient use of available memory bandwidth alongside execution of the model itself.
Meta specifies that moving selection onto the GPU preserves advertisers' targeting rules. The index must handle conditions across dozens of feature categories and reflect real-time changes as ads are created, modified or expire. Faster evaluation operates within these eligibility conditions.
The figures describe separate components and periods
In production, the company reports more than one billion ads evaluated per second for the complete retrieval pipeline on a GPU server. It presents this as an approximately 100-fold increase over the 2024 Andromeda system. The figure measures processing capacity rather than the number of people seeing advertisements.
Separately, Meta describes more than ten billion evaluations per second for the kernel handling user–ad interactions on a GH200 GPU. Index matching exceeds 100 billion per second in the reported tests. These values concern different components and are not equivalent to the throughput of the entire production system.
Meta attributes a 2.4% increase in Facebook conversions to Andromeda developments since 2025. This is a company-reported value for its system, covering a broader reference period than the October announcement. The technical document presents it separately from the speed measurements for the new architecture.
During the NVIDIA collaboration, the team measured a 5.3-fold acceleration of the interaction-scoring kernel, from 2.4 billion to more than ten billion ads per second. According to Meta, that component reached approximately 85% of peak GPU memory bandwidth. The result describes the efficiency of the particular technical operation being examined.
The company also tested representing numerical values with eight bits instead of the 16-bit version. Meta reports an additional throughput increase of up to two times, with accuracy comparable to 16-bit precision. For the complete pipeline, optimized test configurations reached speeds several times higher than current production throughput.
The optimizations were also developed through a collaboration with NVIDIA that began in early 2026. Meta says work continues on Grace Blackwell platforms and on the balance between CPU and GPU resources. These are infrastructure development directions, while the update presented describes advertising retrieval already used in production.
