360Brew: The AI Brain Behind LinkedIn’s New Algorithm

On March 12, 2026, LinkedIn unveiled the most significant overhaul of its feed since the platform was created. At the core of the change: 360Brew, a 150-billion-parameter AI model that replaces the fragmented recommendation machinery built over the years.

That said, it isn't bad news for everyone. Creators who understand how 360Brew thinks are actually seeing their authority grow. Here's what you need to know to play by the new rules instead of against them.

What exactly is 360Brew?

360Brew is a decoder-only foundation model, developed by LinkedIn's FAIT (Foundation AI Technologies) team and publicly detailed in a research paper published on arXiv in January 2025 by Hamed Firooz and his colleagues.

Three technical characteristics to remember:

• 150 billion parameters, trained exclusively on LinkedIn's proprietary data (profiles, posts, professional interactions, job descriptions).
• Architecture derived from Meta's LLaMA 3 family, adapted and fine-tuned by LinkedIn.
• Capable of handling over 30 predictive tasks: feed ranking, job recommendations, connection suggestions, ad targeting, and more.

Before 360Brew, the feed ran on five separate retrieval pipelines working in parallel — trending content, collaborative filtering, geographic trending, industry-specific modules, embedding-based systems — each with its own infrastructure. No single team could optimize across all of them coherently. 360Brew unifies everything into one brain that actually reads the content instead of relying on indirect signals like hashtags or clicks.

A simple way to understand it: imagine 360Brew works like ChatGPT, but internal to LinkedIn. Every time you open the app, the system fires a prompt to its model, something like:

• When you run a search: "Based on this user's profile and the criteria they just entered, give me the list of profiles or content most relevant to display."
• When you open your feed: "Based on this user's profile, interests, engagement history, proximity to authors, and several other criteria, show me the list of posts most relevant for them."

The system holds your data, preferences, and history in memory, combines it with its semantic understanding of content, and composes your feeds, lists, and suggestions from there.

How does the new LinkedIn algorithm work since March 12, 2026?

The system officially announced by Hristo Danchev, Senior Staff TPM at LinkedIn, works in two distinct stages.

Stage 1: Retrieval

A first model — a Causal LLM built on LLaMA 3 — converts each post and user profile into a vector representation in a shared semantic space. In concrete terms, it turns words into numbers that capture their professional meaning.

When you open LinkedIn, this model compares your "semantic fingerprint" (what your profile and interactions say about your professional interests) with that of millions of posts, and narrows the pool down to roughly 2,000 candidates in a few milliseconds. The matching is done using cosine similarity — a mathematical measure of semantic proximity.

This is where cold start becomes a strength. A new user who lists "electrical engineer" in their profile can receive relevant content about power grid optimization or small modular nuclear reactors from day one, with no engagement history. The LLM infers latent interest from meaning, not from history.

Stage 2: Ranking

The 2,000 shortlisted posts then pass through the Generative Recommender (GR), a transformer model that analyzes more than 1,000 of your past interactions as a chronological sequence — not as a bag of independent events.

This distinction is critical. The old system scored each post in isolation. The new one understands your trajectory: if you read three articles about B2B marketing this week, then looked up growth hacking, then saved a post about lead generation, it infers where your interest is heading — not just where it has been.

What has actually changed in your reach?

The numbers tell a brutal but consistent story.

Sources: Algorithm Insights 2026 — Richard van der Blom and cross-analysis of Buffer, Metricool, Socialinsider, and Agorapulse by Xavier Degraux.

What 360Brew rewards (and what it penalizes)

Because it reads content instead of counting reactions, 360Brew radically changes what works.

What's rewarded

• Topical consistency. Publishing regularly on 3 or 4 specific topics sends a clear signal to the model. Your profile becomes a workable "semantic fingerprint."
• Original expertise. The model can evaluate the semantic novelty of a post. An original take, first-party data, or a specific experience performs better than recycled advice.
• Substantive comments. A post that generates three thoughtful comments beats one with thirty likes. The algorithm weights active engagement (comments, shares, DMs) far more than passive engagement.
• Saves and dwell time. When someone saves your post or spends time reading it, 360Brew reads this as a strong utility signal — and widens distribution.
• Profile–content alignment. A SaaS CMO who talks about go-to-market, product-market fit, and B2B growth sends a clean signal. A SaaS CMO who alternates between management, recipes, and meditation sends noise.

What's penalized

• Engagement pods and coordinated tactics (we'll dive deeper into this below — big topic).
• Mechanical engagement bait. "Comment YES if you agree," "Tag someone who needs to see this," "Like if you've been there" — all detected, all suppressed.
• Generic AI-generated content. Posts following predictable patterns without an original perspective are actively deprioritized.
• Editorial inconsistency. Switching topics every post isn't seen as variety anymore — it reads as a lack of positioning.
• Posting too often. Daily posting = -45% reach. The sweet spot is 2 to 3 posts per week.

Why LinkedIn "cheaters" are about to lose everything with 360Brew

This is probably the most underrated shift. For years, a parallel economy of engagement tactics has been built on LinkedIn: pods, comment bots, copy-pasted hooks, clickbait polls, generic comments at scale. All of it is collapsing. Here's why.

How to adapt your LinkedIn content strategy in 2026

Here are the concrete adjustments to make this week, not in six months.

FAQ

Last updated: April 21, 2026. This article will be updated as soon as LinkedIn publishes new technical information on 360Brew or its successors.