LinkedIn's feed runs on one 150-billion-parameter model

LinkedIn replaced the stack of separate ranking models behind its feed with a single 150-billion-parameter language model, and it is now in production. That is why the tactics built on metadata stopped working.

The engineering is public. LinkedIn published 360Brew in January 2025, a decoder-only foundation model of 150B parameters designed to handle ranking and recommendation across surfaces that previously each had their own task-specific model: feed, people, jobs, ads. On 12 March 2026 LinkedIn confirmed the production rollout, describing generative recommenders paired with large language models, running on a GPU-backed ranking architecture.

Almost every article written about the LinkedIn algorithm in the last year has not connected the research paper to the product announcement. It is the biggest change to the platform's distribution in a decade.

What actually changed underneath

The old approach predicted engagement from features. Did this user engage with this author before, does the post carry tags matching their declared interests, what is the historical click rate on this format, and so on. Metadata in, probability out. It worked, and it could be gamed by supplying better metadata.

The new approach reads the content and reasons about it semantically. LinkedIn's own example, in its own words, is a member interested in electrical engineering who engages heavily with content about small modular reactors. A keyword system misses that connection because the terms do not overlap. A language model does not, because it understands that one sits inside the other.

LinkedIn also said the system now surfaces expert commentary on breaking news within minutes rather than hours. That is a comprehension change, not a speed change. The model can tell that a post is informed commentary on an event, without waiting for engagement data to prove it.

The tactics this kills

Hashtags as a distribution lever. They were always weaker than people believed. They are now close to irrelevant, because the model reads what your post is about and does not need you to tag it. Two or three at the end for human navigation is fine. Ten is noise that makes the post look like marketing.

Keyword stuffing and topic tagging. Writing manufacturing, supply chain, logistics, ERP into a post to catch multiple audiences no longer catches anything. The model reads the actual content and forms its own view.

Engagement pods and the mechanical opener. Bait phrasing still generates the initial interaction, but the model now evaluates the post itself. A post engineered for a reaction and containing nothing has a shorter runway than it did.

Broad safe posting. This is the important one. Generic content used to work because it matched broad interest categories. Semantic ranking rewards content that is precisely about something, because precision is what lets the model find the right small audience with genuine interest.

What this rewards instead

Specificity that a model can locate in semantic space. Not a sentence saying we improved our procurement process. Instead: which category, what the lead time was before and after, what the constraint was, what you changed, what it cost. That post can be matched to people who care about that exact problem, even if they have never used your words for it.

Real domain vocabulary. Standards, methods, materials, systems, named processes. This used to be discouraged as jargon that limits reach. It is now the signal that tells the model exactly where your post belongs.

Commentary with a position. An informed take on something that just happened, from someone who does the work, is now identifiable as that within minutes of posting.

Depth over breadth in topic selection. If your company posts about six unrelated things, the model has a harder time forming a coherent view of who should see you. A consistent topical identity is worth more than it was.

What to do differently on Monday

Rewrite your posting brief around one rule: could a reader tell which company wrote this without the logo? If not, the model cannot place it either.

Take your last twenty posts and mark each one for whether it contains a number from your own operations, a named method or standard, or a specific decision with an alternative rejected. Whatever percentage you get, that is roughly the proportion of your output the new system can do anything useful with.

Cut hashtag counts to three. Reallocate the effort to the first two lines, which still determine whether anyone reads far enough for the ranking to matter.

Let your technical people post in their own vocabulary. The instinct to simplify everything for a general audience was correct for a system that matched broad categories. It is counterproductive for one that reasons semantically.

The honest caveats

LinkedIn has not published per-post ranking weights and will not. What is documented is the architecture and the reasoning capability, not a formula. Anyone giving you exact weightings for the 2026 feed is making them up.

This also does not repeal the basics. Dwell time and comments still matter, posting into an empty network still fails, and a semantically precise post about something nobody cares about still gets nothing.

And it makes some things harder. If your business genuinely sells something generic, semantic ranking does not help you, because there is no precise thing for the model to attach you to. That is a positioning problem the feed has now made visible.

Write like someone who does the work, using the words they use. That is not a style preference any more. It is how the ranking system finds you.