LinkedIn 360Brew: what it is, and what it is not

360Brew is a LinkedIn research model. It is not, on the evidence of any primary source we could find, the system that decides who sees your posts. LinkedIn published a long technical description of its rebuilt feed on 12 March 2026, and 360Brew does not appear in it once. The 360Brew paper itself has been removed from arXiv by administrators. Somewhere between those two facts, the LinkedIn advice industry decided that a 150-billion-parameter model named 360Brew now ranks your content, and started publishing strategy guides on that basis.

What is 360Brew?

360Brew V1.0 is described in a paper credited to a large LinkedIn author list, including Hamed Firooz, Maziar Sanjabi and Adrian Englhardt, first posted to arXiv on 27 January 2025. The idea is genuinely interesting. Recommendation systems have historically been built as collections of specialised models, each predicting one thing: the chance you click, the chance you comment, how long you are likely to linger. Each model needs its own features, its own training pipeline and its own team. The paper's argument is that a single large decoder-only model, given the problem in text form, can do all of it at once.

The specifics worth holding on to are these. The model is 150 billion parameters. It covers more than 30 predictive tasks across LinkedIn surfaces. It works through a text interface instead of hand-built numeric features. And the headline claim, in the paper's own words, is performance “comparable to or exceeding those of current production systems based on offline metrics, without task-specific fine-tuning”.

Read that sentence slowly, because the entire misunderstanding lives inside it. “Comparable to or exceeding current production systems” is a statement about a benchmark comparison. “Based on offline metrics” means the comparison was run against logged data rather than against live traffic. A model can beat a production system on offline metrics and never ship, and this happens constantly, because offline gains routinely fail to survive contact with a real serving budget and real users.

Why does the withdrawal matter?

Every version of the paper on arXiv (v1 through v4, the last posted 23 August 2025) now carries an administrative notice. The wording is specific:

This version has been removed by arXiv administrators as the submitter did not have the right to agree to the license at the time of submission.

Be careful about what this does and does not mean. It is not a retraction. Nobody has said the results are wrong, and there is no fraud allegation here. It is a permissions problem: whoever uploaded it was not authorised to license it for distribution, which is an ordinary thing to happen when an employee posts corporate research.

What it does mean is that the single most-cited artefact in this entire conversation is no longer a publicly readable document. If you are writing strategy advice that depends on the contents of that paper, you are now depending on other people's summaries of a paper they can no longer link you to. That is the point at which a claim stops being verifiable and starts being folklore, which is a pattern this niche repeats often enough that we gave it its own section in our algorithm guide.

What did LinkedIn actually ship in 2026?

On 12 March 2026, LinkedIn published “Engineering the next generation of LinkedIn's Feed” by Hristo Danchev on its engineering blog. This is the primary source for what changed, and it is a substantially different system from the one described in the 360Brew paper.

The feed was rebuilt in two halves. The first half is retrieval: choosing a few hundred candidate posts out of everything that exists. LinkedIn replaced a multi-source retrieval system with a single unified one built on an LLM-based dual encoder. One encoder builds a representation of you from profile information, skills, work history, education and the ordered sequence of posts you have engaged with. The other builds a representation of candidate content. Posts are retrieved by similarity between the two, which is why the system can pull something relevant from outside your network without anyone having shared it to you.

The second half is ranking: ordering those candidates. LinkedIn replaced independent post-by-post scoring with a Generative Recommender, a transformer that treats your feed history as a sequence and reads more than a thousand of your past interactions in order, with a Multi-gate Mixture-of-Experts prediction head over DCNv2 experts. The difference from the old approach is that a post is no longer scored in isolation. It is scored in the context of the sequence of things you have already responded to.

The post also reports engineering figures: percentile bucketing of numerical features improved Recall@10 by 15%, adding one easy and two hard negatives during training improved recall by a further 3.6%, retrieval latency sits under 50ms, and custom attention kernels delivered a 2x speedup, against a platform serving more than 1.3 billion members.

Those numbers are about the model's own performance, not about your reach. No part of that post promises anyone more impressions. It is worth saying plainly because several summaries have converted “Recall@10 improved 15%” into implied audience growth, which is not what the metric measures.

Retrieval and ranking are different jobs, and the distinction matters

Most feed explainers collapse these into one thing called “the algorithm”, which is why so much advice is confused. They are separate problems with separate failure modes.

Retrieval is a search problem run against an index of essentially everything available, and it has a hard latency budget: LinkedIn reports sub-50ms. It cannot afford to think carefully about any individual post. Its job is to not miss things. Ranking is a much more expensive computation applied to a much smaller set, and its job is to order what retrieval already found.

The practical consequence is that these two stages fail differently, and knowing which one failed tells you what to fix. If your post never enters anyone's candidate pool, no amount of hook rewriting will help, because nobody got far enough to read the hook. That is a retrieval outcome, and it is driven by whether your post is about something the system can recognise and match to a reader. If your post is retrieved and then buried, that is a ranking outcome, and it is driven by predicted response in the context of that reader's history. Advice that ignores this distinction will tell you to fix your first line when the actual problem is that you wrote about six unrelated things in one post.

360Brew versus what shipped: a side-by-side

These could still be related pieces of work. Large companies often publish research that later informs a shipped system under a different name, and LinkedIn is under no obligation to tell anyone which internal model does what. The honest position is the narrow one: no public source connects them. Anyone telling you that 360Brew reads your posts is describing something they cannot show you.

Where did the LLaMA 3 claim come from?

A specific detail worth isolating, because it repeats everywhere: many articles state that 360Brew is built on LLaMA 3. The paper abstract does not name a base model. We could not trace that claim to any LinkedIn source. It may well be true, and it may have appeared in the full paper text that is no longer retrievable, but as of this writing it is a claim circulating without a citation anyone can check.

This matters less for its own sake than as a test. If a page tells you 360Brew is a LLaMA 3 derivative that ranks your feed, and links only to another blog, you have learned something useful about how carefully the rest of that page was researched.

What actually changes for the person writing posts?

Less than the coverage implies, and in a different direction than most of it suggests. Three things follow from retrieval-by-meaning rather than retrieval-by-network.

Your audience is less fixed than it was

When candidate posts are pulled by embedding similarity against a reader's interests, a post can reach someone with no connection path to you. That cuts both ways: a post about a recognisable subject can travel further than your follower count implies, and a post about nothing in particular has fewer places to land, because there is no coherent interest to match it against. If your reach has been moving unpredictably, our guide on why LinkedIn reach drops covers the mundane explanations first, which are still the more likely ones.

Topical coherence is now a mechanism, not just advice

“Pick a lane” used to be soft positioning advice. Under a system that builds a vector representation of what you write about, it has a mechanical reading: a body of work about one identifiable subject produces a cleaner signal than the same volume scattered across six unrelated ones. This is the strongest practical implication of the rewrite, and it is the argument for content pillarsthat does not depend on anyone's opinion.

Sequence-aware ranking makes single posts a bad unit of measurement

If the ranker reads a thousand of a reader's prior interactions as an ordered sequence, then whether your post surfaces for a given person depends heavily on what that person did last week. Post-level variance was already high. This makes judging a post by its first hour even less defensible than it was, which is the same conclusion our golden hour analysis reached from a completely different direction.

How do you write for retrieval by meaning?

This is the one section of this article with genuinely new practical content, so it is worth being concrete rather than gesturing at “quality”. If candidate selection runs on learned representations of what a post is about, then a post needs to be about something in a way a machine can represent. Four habits follow.

Put the subject in the text, not in your head

A post that opens “This changed everything for me” and only reveals its subject in paragraph four is hard to represent, because the first thing the system sees is content-free. This is not a new rule dressed up in machine-learning language. It is the same reason a human scrolls past. Our hooks guidegets to the same place from the reader's side.

Use the words your reader would use

Embedding-based retrieval is more forgiving of synonyms than keyword matching, so you do not need to repeat an exact phrase. But it still represents what is on the page. If you write about “the thing we built” instead of naming the category, you have removed the signal without gaining anything. Write the plain noun.

One post, one subject

A post covering hiring, fundraising and a conference recap has a representation that sits between three interests and matches none of them well. Split it. This costs you nothing, because you needed three posts anyway.

Let your body of work be consistent, not identical

Consistency of subject is different from repetition of format. Writing about the same domain from different angles builds a clear picture of what you cover. Posting the same template daily does not, and it gives readers a reason to stop responding, which the ranking half of the system will notice. The practical version of this is a small set of recurring themes, which is what our content strategy guide lays out in more detail.

Notice that none of these are tricks, and none of them depend on the rewrite being described accurately by anyone. They are what you would do anyway to be legible to a human reader. That is the useful test for any algorithm advice: if it only makes sense as a manipulation of a ranker, it will expire.

Does this explain the reach changes people reported in 2026?

Partly, and with more humility than most of the coverage shows. A retrieval rewrite redistributes attention. When the candidate pool for every reader is rebuilt on a different basis, some accounts gain and some lose, and the losers write posts about it while the winners quietly enjoy it. That asymmetry alone guarantees that any change of this size reads publicly as a decline.

What we cannot tell you, and what nobody outside LinkedIn can, is the size or direction of the effect for any particular account. LinkedIn published no before-and-after reach distribution. Every “reach fell 40% after the update” figure in circulation comes from someone's own account or a small self-selected panel, which cannot separate a platform change from seasonality, a change in their own posting, or ordinary variance. Our guide to organic reach decline works through why these panels almost always overstate the effect.

The honest summary: a large change happened, it plausibly moved reach around, and anyone quoting a precise number for what it did to accounts in general is making it up.

What does our own data say, and what does it no longer cover?

We analysed 12,988 LinkedIn posts from 65 creators, comparing the top decile against the bottom half. The patterns we found are about what makes a human stop and respond, and those hold up reasonably well across ranking changes: first-person openers appear in 19.6% of top-decile posts against 10.1% of the bottom half, native video appears in 26.3% against 10.1%, and the comment gap is far wider than the reaction gap (a median of 72 comments in the top decile against 5 in the bottom half).

Here is the honest limitation, and it is a real one: that dataset was collected before the March 2026 feed rewrite. It describes what earned engagement under the previous distribution system. Reader psychology does not reset when an architecture changes, so we expect the hook and format findings to age reasonably. Anything in our data that is really a claim about distribution mechanics should be read as historical. We would rather say that than quietly present pre-rewrite data as a description of the current system, which is what most of the pages citing old studies are doing right now.

Full method, cohort rules and caveats are in the engagement study.

The documented timeline

Almost nobody writing about this lays out the dates, which is how a 2025 research posting and a 2026 product change ended up fused into one event. The public record is short enough to list in full.

Thirteen and a half months separate the paper from the product announcement, and the two documents describe different architectures. Any article presenting them as one announcement has not read both.

What do we still not know?

Listing the gaps is more useful than filling them with guesses, and it is the part every other page on this subject skips.

• Whether 360Brew powers anything user-facing. No public source confirms or denies it.
• What the ranker optimises for. LinkedIn describes the architecture, not the objective. Which predicted actions carry the most weight is not published, so every weighting figure you have seen is invented.
• How much distribution now comes from outside your network. The system can do it. No figure has been published for how often it does.
• Whether any of this changed reach on average. No before-and-after data exists publicly.
• How quickly the representation of you updates. If you change what you write about, nobody outside LinkedIn knows whether the effect takes days or months.

Anyone who answers those five questions confidently is guessing. That is a reasonable filter to apply to the next algorithm article you read.

What should you stop doing?

• Stop rewriting your strategy every time an architecture is announced. The retrieval stack changed. What makes a person read to the end did not.
• Stop treating secondary summaries as sources. The two primary documents here are one engineering blog post and one withdrawn paper. Most articles on this subject cite neither.
• Stop expecting a fixed number.Claims of the form “360Brew weights comments 4x” are invented. Nothing in the public record assigns weights to anything.
• Stop optimising for retired mechanics. Several popular tactics predate not just this rewrite but the one before it. Our notes on hashtags and external links trace where those rules came from and how much evidence sits behind them.

How do you check an algorithm claim yourself?

This is the durable skill, and it takes about ninety seconds.

1. Find the link. If the claim has no link, it has no source. Stop there.
2. Follow it one hop. Most links in this niche go to another blog making the same claim. Keep going until you reach a primary document or a dead end.
3. Check who published it.LinkedIn's engineering blog, LinkedIn Help, an arXiv paper or a named LinkedIn employee speaking on the record are primary. A tool vendor's content marketing is not.
4. Check the date, then check it again. A 2025 algorithm explainer describes a system that no longer exists in the same form.
5. Separate measurement from mechanism.“Posts with video get more engagement” is a measurement. “The algorithm boosts video” is a mechanism claim, and it needs much stronger evidence.

The short answer, one more time

360Brew is a real and reasonably ambitious piece of LinkedIn research whose paper is no longer publicly available. The feed you post into was rebuilt in March 2026 around an LLM dual encoder and a sequential transformer ranker, described by LinkedIn itself, with no mention of 360Brew. Anyone selling you tactics tuned to 360Brew specifically is working from a document they cannot show you, about a system nobody has confirmed is running.

The part that survives all of this is unglamorous. Write about something specific enough to be recognisable. Say it in a way that earns the second line. Give people a reason to reply rather than scroll. That was true under the old retrieval stack, it is true under this one, and it will be true under the next one.

Sources

• Hristo Danchev, “Engineering the next generation of LinkedIn's Feed”, LinkedIn Engineering Blog, 12 March 2026.
• 360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation, arXiv 2501.16450. All versions removed by arXiv administrators; withdrawal notice quoted above.
• Our analysis of 12,988 posts from 65 creators, described in the engagement study.