360Brew is the name LinkedIn has given to a large, decoder-only foundation model built to handle ranking and personalization work, feed, jobs, ads, search and similar surfaces, with one general-purpose model instead of dozens of separate, narrow models. LinkedIn has described it in its own engineering research as a shift from many specialized systems toward a single model that can be adapted across tasks.
For creators and B2B marketers, the practical takeaway is not a secret setting to flip. It is that a ranking system built to actually reason about language and context is likely more forgiving of genuine, clear writing and less forgiving of keyword games built to exploit older, narrower signals.
TL;DR
360Brew is LinkedIn's name for a large, decoder-only foundation model applied to ranking and personalization tasks across surfaces like the feed, jobs, ads, and search, replacing (or supplementing) many separate specialized models with one general-purpose system. It is not a public chatbot. It does not hand creators a new trick. It rewards the same fundamentals that have always mattered: clear writing, consistent topics, and genuine engagement, and is plausibly less forgiving of keyword-stuffing style tactics built around older, narrower ranking signals.
A side-by-side view of the general approach before and after moving to a single foundation model, based on how LinkedIn's own engineering research has described it.
| Aspect | Before (specialized models) | After (360Brew foundation model) |
|---|---|---|
| Ranking approach | Dozens of narrow, specialized models, each trained for one job (feed ranking, ads, jobs, search). | One large, decoder-only foundation model that can be adapted to many ranking and recommendation tasks. |
| How signals get combined | Hand-engineered features feed into task-specific scoring models built and tuned separately. | The model reads profile, activity, and content as natural-language style context and reasons across it. |
| Adding a new surface (jobs, notifications, ads) | Usually means training and maintaining another dedicated model from scratch. | The same base model can, in principle, be prompted or fine-tuned toward a new surface rather than rebuilt. |
| Understanding content | Relies heavily on keyword, engagement-history, and metadata signals to infer topic and quality. | A foundation model trained on language can reason about what a post actually says, not just its metadata. |
| Cold start (new posts, new creators) | Struggles until enough engagement history accumulates for the specialized model to trust a signal. | General language understanding can make a first-pass judgment about quality before history builds up. |
| Explainability for creators | Opaque, and so are most modern ranking systems, but at least conceptually simple: separate scores per surface. | Still opaque from the outside. One larger model can be harder, not easier, to reason about intuitively. |
The vocabulary that shows up whenever 360Brew gets discussed, defined without the jargon.
Foundation model
A single large model trained on broad data that can be adapted to many different downstream tasks, instead of training one narrow model per task.
Decoder-only architecture
The same general architecture family used by most modern large language models. Applying it to ranking means treating recommendation as a language-style prediction problem.
Ranking model
The system that decides the order content appears in a feed, search result, or job list, after a broader retrieval step has already pulled a candidate pool together.
Personalization signal
Any input, such as your activity history, connections, or stated interests, that a model uses to tailor what it shows a specific member.
Retrieval vs. ranking
Retrieval narrows millions of possible posts down to a manageable candidate set. Ranking then orders that smaller set for a specific viewer. Foundation models are typically applied at the ranking stage.
Distribution stages
The idea that a post is shown to a small seed audience first, then expanded to a wider pool if early engagement quality is strong. This general concept predates 360Brew and still applies in an AI-driven ranking system.
Lifast helps you write clear, consistent LinkedIn posts on your real expertise, the kind of content a language-capable ranking system has an easier time recognizing as genuinely valuable.
Try Lifast Free90 days of consistent posting. No ads.
A short explainer covering the same ground in video form, useful if you would rather watch than read.
"A decoder-only foundation model for personalized ranking and recommendation," is how LinkedIn's own engineering team has described 360Brew in its published technical writing about the project.
Read LinkedIn Engineering's post on 360BrewThe underlying research paper lays out the technical case for consolidating many specialized ranking systems into one foundation model that can be adapted across tasks.
View the 360Brew research paper on arXivFor ongoing technical writing from the team building LinkedIn's ranking, recommendation, and AI infrastructure more broadly, LinkedIn maintains a public engineering blog covering this and related work.
Browse the LinkedIn Engineering blogDo
Write posts with a clear, coherent point a language-capable model can actually parse and understand.
Keep a consistent topical identity so the model can build a confident picture of what you post about.
Let comments and replies read as genuine conversation, not generic filler.
Vary your formats (text, carousel, native video) since a broader model is built to reason across formats, not just one.
Give the model time. Judge changes to your results over months of posting, not a handful of posts.
Don't
Do not assume hashtag stuffing or keyword repetition will trick a model built to understand actual meaning.
Do not treat this as a one-time trick to learn and exploit. Foundation models are retrained and iterated on.
Do not abandon the fundamentals (a real hook, a clear idea, a reason to comment) that still matter regardless of the ranking system underneath.
Do not assume it is only about jobs or ads. LinkedIn has described this class of model as applicable across ranking and personalization work broadly, including the feed.
Do not panic-rewrite your whole content strategy on a single week of unusual numbers. Isolate whether it is the model, a seasonal dip, or something else first.
None of this requires knowing 360Brew's internals. It requires treating a language-capable ranking system as a reason to lean into fundamentals, not away from them.
Write for a reader that understands meaning, not just metadata
A foundation model trained on language is built to reason about what a post is actually saying. That rewards clarity, a real point, and honest writing over keyword-stuffed or template-shaped text designed to game older, narrower signals.
Build a consistent, recognizable topical identity
Systems built to reason over broad context tend to reward creators whose posting history forms a coherent picture (what you talk about, who you help, what your expertise is) rather than accounts that jump between unrelated topics every week.
Treat every post as training signal, not a one-off bet
Each post you publish, and how people genuinely respond to it, becomes part of the history a ranking system reads when it decides how to treat your next post. Consistency compounds more than any single viral swing.
Prioritize real engagement over engagement bait
Generic prompts like 'agree?' or comment-for-comment exchanges are the kind of pattern a model with broader context is well positioned to recognize as low-value, formulaic engagement rather than genuine interest.
Keep publishing across multiple formats
A single foundation model that can reason across surfaces is a reasonable bet on being able to evaluate text, carousels, and video with more nuance than three separate narrow systems each doing their own thing. Do not put all your output into one format.
Do not chase leaked 'hacks' as permanent rules
Any specific trick that spreads widely as 'the 360Brew hack' is, almost by definition, a pattern a model that understands language and context is well-positioned to eventually discount. Durable principles (clarity, consistency, genuine value) age better than tactics.
Watch your own results, not just what the internet says
LinkedIn has not published a public playbook for 360Brew specifically. Treat aggregate creator commentary as directional, and pay closer attention to what actually happens to your own reach and engagement over a meaningful sample of posts.
Use tools that help you post consistently and cleanly, not to game the system
Consistency, clarity, and genuine expertise are the inputs that hold up regardless of which model is doing the ranking underneath. That is a better bet than any specific workaround aimed at a system LinkedIn has not fully detailed publicly.
Keep your profile and content aligned
A model reasoning holistically about a member is plausibly better positioned to notice when your headline, about section, and recent posts tell three different stories. Keep your presence coherent end to end.
Give any algorithm shift a real evaluation window
Foundation models get retrained and refined over time. A dip or bump in the first few weeks after a change like this is reported is not reliable evidence of a permanent new rule. Judge over a quarter, not a week.
None of the shift toward foundation-model ranking changes the basic job of a good LinkedIn post: say one clear thing well, in your own voice, about something you actually know. Tools like Lifast can help you turn what you already know into consistent, clearly written posts, which is exactly the kind of content a ranking system built to understand language is best positioned to recognize and reward, rather than trying to reverse-engineer a model LinkedIn has not published a public playbook for.
"360Brew is a single algorithm change I can crack."
It is a shift in the underlying model architecture behind multiple ranking and recommendation systems, not a single new rule with a workaround. There is no confirmed trick that reliably games it.
"This means LinkedIn is running my posts through a public chatbot."
360Brew shares an architecture family with large language models, but it is purpose-built internal ranking infrastructure, not a consumer-facing chat product.
"Hashtags and keywords matter more now."
The opposite is more plausible. A model built to understand actual meaning has less reason to lean on surface-level keyword signals than older, narrower systems did.
"This only affects jobs and ads, not the feed."
LinkedIn's own engineering research describes the approach as applicable broadly across ranking and personalization tasks, which is a wider scope than any single surface.
"Nothing about my content strategy needs to change."
The fundamentals stay the same, but tactics that only ever worked by exploiting narrow, older signals are a reasonable thing to retire regardless of whether 360Brew specifically is what catches them.
Hypothetical, plausible scenarios meant to illustrate the general dynamic, not documented case studies of any specific account.
A B2B founder had built a habit of packing posts with trending hashtags and industry buzzwords, on the theory that more keywords meant more matches with older, narrower ranking signals. Under a ranking system built to actually parse meaning, that habit stopped being useful, and posts written more plainly, with one clear point, started performing more consistently for the same account.
A consultant posted about five unrelated topics across a month: fitness, parenting, sales tips, a political opinion, and a product update. A system that reasons over a member's full posting history has less coherent context to work with than it would for a creator who stays in one recognizable lane, which can mean less confident targeting of the right audience for any single post.
A recruiter posted three times a week on one recognizable theme (hiring in a specific niche) for six months straight, replying to comments in plain language rather than copy-paste responses. Nothing about that habit depends on knowing anything about 360Brew specifically. It is simply the kind of consistent, legible signal a language-capable ranking system has more to work with when deciding who to show to whom.
A marketer wrote a single unusually well-performing post, then assumed they had 'found the 360Brew hack' and repeated the exact same structure for the next ten posts. Performance regressed toward the account's normal baseline within two weeks, which is the expected pattern when a format novelty wears off rather than evidence any specific trick was ever a durable rule.
LinkedIn's engineering research describes the model as applicable broadly across ranking and personalization tasks. It has not published a granular, surface-by- surface breakdown of exactly where it is deployed at any given moment, and that detail can change as the system evolves.
Maybe, but it is one of many possible explanations. Seasonality, changes to your own posting habits, and normal week-to-week variance in reach are all at least as common. Attribute a change to a specific named system only if it holds up across a large, sustained sample of your own posts.
The safest adjustment is to double down on writing that sounds like the client's actual voice and covers the client's actual expertise clearly, rather than leaning on generic, templated structures built to satisfy older, narrower ranking signals.
Running dozens of separate, specialized ranking models (one for the feed, another for jobs, another for ads, another for search, and so on) is expensive to maintain. Every model needs its own training data pipeline, its own feature engineering, and its own team keeping it current. A single foundation model that can be adapted across those surfaces is, at least in principle, a way to consolidate that engineering overhead while sharing improvements across every surface at once.
The tradeoff is that a single large model is a bigger, more complex system to reason about from the outside. Where a specialized model has a narrower, more predictable failure mode, a foundation model's behavior across many tasks is harder for anyone, including LinkedIn's own engineers in some cases, to fully anticipate ahead of time. That is part of why creators should treat specific tactical claims about it with some skepticism.
It does not mean LinkedIn quietly started running your posts through a public chatbot. 360Brew is a purpose-built ranking and recommendation system, described in LinkedIn's own technical writing as decoder-only architecture applied to personalization tasks, not a consumer chat product. The similarity in underlying architecture family does not mean the same product experience.
It also does not mean the fundamentals of good LinkedIn writing changed overnight. A clear hook, one focused idea, a reason to comment, and consistent posting still describe what tends to perform well. What a shift like this can change is how forgiving or unforgiving the system is toward keyword games and engagement-bait patterns that were built to exploit older, narrower signals.
360Brew is one visible example of a broader trend across large platforms: replacing many narrow, specialized ranking models with fewer, larger, more general ones. Expect this pattern to continue and to show up under different names on other platforms over time. The specific model name matters less than the direction it points in.
The most durable response to any of these shifts is the same: write content that is genuinely clear and useful to a human reader, stay consistent in what you post about, and treat engagement as something you earn rather than something you engineer around a specific system's quirks. That approach ages well regardless of which model happens to be doing the ranking in a given year.
Straight answers to what creators and marketers actually ask once they hear the name 360Brew.
360Brew is the name LinkedIn has given to a large, decoder-only foundation model built to handle ranking and personalization tasks (feed, jobs, ads, search and similar surfaces) with one general-purpose model, instead of a separate specialized model for each surface. It is infrastructure, not a public-facing product you interact with directly.
It shares the same broad architecture family (decoder-only models, the kind used across much of the current generation of large language models), but it is built and applied to LinkedIn's internal ranking and recommendation problems, not as a consumer chat assistant. Sharing an architecture family is not the same as sharing a product.
It is more accurate to say it changes the underlying machinery behind ranking and recommendation, rather than introducing a single new named 'algorithm' the way creators often talk about it. The general concepts that still matter, such as early engagement quality and a coherent posting history, remain relevant even as the model doing the evaluating changes.
There is no publicly confirmed answer either way. A model built to reason about content meaning rather than relying heavily on existing follower counts and engagement history could, in principle, help newer or smaller accounts get a fairer first look. Treat that as a plausible direction rather than a guarantee until it is borne out by broad creator data over time.
Not dramatically. The fundamentals, a clear point, genuine value, consistent posting, and real engagement, remain the best bet regardless of the ranking system underneath. What is worth dropping is any tactic that only ever worked by exploiting a narrow, keyword-based signal, since a more language-capable system is better positioned to see through those patterns.
LinkedIn's engineering team has published technical writing describing 360Brew as a decoder-only foundation model for personalized ranking and recommendation. That is the most direct primary source, and it is linked in the source cards on this page, alongside the associated research paper.