LinkedIn's decoder-only foundation model for feed ranking, which judges a post's relevance from its text rather than from its engagement counts alone.
360Brew is the large language-model-based recommendation system LinkedIn detailed in a 2025 research paper: a 150B-parameter, decoder-only model. Instead of the older multi-stage pipeline that leaned on numeric IDs and hand-built features, it reads the actual content of a post and a member's history as text and predicts relevance directly.
The practical consequence for anyone posting is that the model can reason about what a post says, not just how many reactions it has collected. Generic, low-information, or off-topic writing has fewer places to hide, because relevance and expertise signals are read from the language itself.
It is a ranking and retrieval model, not a moderation system. It does not ban content. It decides how far a post travels by estimating how relevant it is to each potential viewer, which is why two posts with identical formatting can see very different reach.
Frequently asked questions
Is 360Brew real?
Yes. LinkedIn published research describing 360Brew, a decoder-only foundation model for personalized ranking and recommendation, and has described moving feed ranking toward LLM-based systems. The exact production configuration is not public.
Does 360Brew penalize AI-written posts?
Not by detecting AI directly. It scores relevance and expertise from the text. AI drafts that are generic or restate common knowledge tend to score lower on those signals, which reads as reduced reach even though nothing is explicitly flagged.