360Brew: the model that decides what you see on LinkedIn

360Brew is a 150-billion-parameter decoder-only foundation model that LinkedIn built to handle ranking and recommendation across the platform. LinkedIn's Foundation AI Technologies team described it in a paper published on arXiv in January 2025, and it is the system most often meant when people talk about the 2026 LinkedIn algorithm.

The headline claim in the paper is consolidation. 360Brew is described as capable of solving over 30 predictive tasks across the platform, tasks that were each conventionally addressed by dedicated models developed and maintained over multiple years. One model, in place of a fleet of them.

It is built on a Mixtral 8x22 pre-trained mixture-of-experts base and was trained on LinkedIn's first-party data from non-EU members over roughly nine months by a small team.

In two stages. A retrieval stage narrows an enormous pool of eligible posts down to a couple of thousand candidates in a few tens of milliseconds, and then the large model ranks those candidates to produce the order you actually scroll.

The split exists for a boring reason: cost. You cannot run a 150-billion-parameter model over every post on the platform for every member every time they open the app. So a cheap, fast stage throws away almost everything, and the expensive, smart stage only looks at what survived.

Both stages matter to you, and they fail differently. Retrieval decides whether your post is even considered for someone. Ranking decides where it lands once it is.

The system reads. The older ranking stack consumed hundreds of thousands of handcrafted numeric features: counts, rates, categories, engineered signals. The paper describes 360Brew processing all input as text instead, which lets it infer what a post is about rather than being told through metadata.

The paper puts this in engineering terms, saying the approach eliminates the need for feature engineering and for maintaining complex directed acyclic graphs of model dependencies, and describes it as replacing labour-intensive feature engineering with a compute-driven approach.

For a writer, the practical translation is that hashtags, keyword stuffing and format tricks are worth less than they were, because they were mostly ways of feeding metadata to a system that needed metadata. A model reading your sentences does not need you to tell it the topic in a hashtag. It also cannot be fooled by one.

The second change is context. The paper tested context lengths up to 88,000 tokens and reports performance improving as more of a member's history is included. The system is not scoring your post in isolation, it is scoring your post against a long read of what that particular person has engaged with.

The paper reports a larger gap over the previous production system for members with fewer interactions, which is the cold-start case. A model that can read a post and a profile can make a reasonable guess about relevance without needing a long behavioural history to learn from.

That cuts both ways, and it is worth being honest about which way is which. It should mean a new or small account is less penalised simply for being new, because relevance can be inferred from the content rather than only from accumulated engagement.

It also means distribution is narrower and better targeted. A post that would previously have been sprayed at your whole network now goes to the subset the model believes will actually care. That is a better experience for readers and it looks like a reach collapse on your analytics tab. Both are true at once, and it is a large part of why reach numbers fell so sharply in 2026.

Be about something, specifically and repeatedly. If relevance is inferred from language, then the clearer and more consistent your subject matter, the better the model can work out who to show you to. Scattered topics give it nothing to build a picture from.

1. 1.Pick a narrow lane and stay in it long enough for a pattern to exist. Three or four recurring subjects beats twenty one-off takes.
2. 2.Say the specific thing. A model reading for meaning gets more from one concrete number, name or situation than from a paragraph of category-level advice.
3. 3.Drop the hashtag habit. They were metadata for a system that needed metadata. Keep one or two if they genuinely help a human, not because you think they route the post.
4. 4.Write so it makes sense cold. Increasingly your reader is someone the model picked on topical relevance, not someone who knows you.
5. 5.Stop writing hooks that oversell the body. The ranking objective includes whether people stay, so a promise the post does not keep is measurably worse than no hook at all.

The uncomfortable implication is that there is no longer much daylight between writing well and ranking well. A system that reads for meaning rewards posts with meaning in them. That is genuinely harder than learning a format, and it is also the only durable advantage available, because a format everyone can copy stops working the week everyone copies it.

The hard part is rarely knowing what you think, it is getting it onto the page without flattening into the voice every draft defaults to. innernote interviews you the way a good ghostwriter would, learns how you actually talk, and drafts from your own material and opinions rather than from the internet's average take on your topic. Free trial, no card.