1. What LinkedIn’s 360Brew is – in plain-ish English
It is a single foundation model for ranking and recommendations at LinkedIn, designed to handle many predictive tasks that used to require separate systems (feed, people, jobs, ads).
So rather than lots of different models, built and tweaked over many years, trying to join up and work together, 360Brew can be seen as one huge system replacing allof the different models and doing it all.
1.1. Scale & design:
150B-parameters, decoder-only; trained primarily on first-party LinkedIn data for many tasks; the source paper describes it as a pre-production model developed over nine months.
Source: arxiv
1.2 Why this is different:
Prior to this, LinkedIn’s production stack (e.g., LiRank) relied on a multi-stage pipeline with specialised models and heavy feature engineering. 360Brew aims to replace a lot of that hand-built feature work with a model that can understand text and reason across tasks.
Source: di.acm.org
1.3 What this implies for LinkedIn users / creators:
The new system can interpret your words, your bio, company names, skills, and post context more like a person – not just count likes or hashtag use.
2. How this changes visibility, engagement, and reach
2.1 Stronger content understanding → topic clarity wins
The paper emphasises replacing ID/feature-heavy models with a text interface so the model can generalise across surfaces and tasks. For users and creators, clear and specific language about your pillar topics increases the chance the system identifies you as relevant for that topic and audience.
2.2 Profile–post consistency is more important
With a unified model reading both member profiles and content, consistent signals across your headline, about section, experience, and posts help the model match your work to the right readers (audience) – and help you appear in suggested content, people to follow, and beyond.
2.3 Cold-start and niche discovery may improve
360Brew is designed to generalise (including out-of-domain tasks) and reduce the need for hand-crafted features, which typically improves cold-start and long-tail matching. That can help niche experts if their language is clear and entities are named.
2.4 Early feedback and dwell still matter
LinkedIn has publicly discussed dwell time and a two-pass ranking funnel (candidate generation → ranking). Those signals don’t vanish; they feed the model’s understanding of quality and relevance. Meaningful comments and read time remain powerful.
2.5 Network relevance remains a core gate
LinkedIn’s feed continues to weight who you interact with (DMs, comments, saves, follows) and community relevance; 360Brew doesn’t replace social signals – it should help interpret them better.
3. Old vs new: what actually changed?
| Area | Pre-2025 LinkedIn (LiRank era) | 360Brew era (what’s new) |
|---|---|---|
| Architecture | Multi-stage, many task-specific models, heavy feature engineering | Single foundation model with a text interface across many tasks |
| Content understanding | Text used as features for other models | Text is the primary interface; model “reads” posts/profiles and reasons |
| Generalisation | Models tuned per surface (feed, jobs, ads) | Cross-surface generalisation by design |
| Cold-start & long tail | Harder – ID/feature heavy | Better potential matching for niche content and new accounts |
| What helps creators | Network, dwell, topical relevance | Same – plus clear topic pillars, entity names, and consistent expertise |
4. Industry reactions (what practitioners are saying)
Multiple practitioners and researchers have highlighted 360Brew as LinkedIn’s large model for ranking and recommendations, noting the 150B scale and its aim to unify tasks.
Source: Alexander Low
Marketer guides published this year updated their advice to reflect LinkedIn’s LLM-driven feed, calling out clarity, entities, and expertise as practical levers. (Unofficial, but widely circulated.)
Source: trustinsights.ai
Independent explainers continue to remind users that dwell, early feedback, and network relevance remain key operational signals even as the model changes under the hood.
Source: Hootsuite
(Note: reaction posts are commentary, not official policy – but they show how the field is interpreting LinkedIn’s own research.)
5. What to change in your posting – today
5.1 Tighten your topic pillars
Pick 2–4 themes you want to be known for. Name people, companies, tools, and skills in natural language – don’t write around them. This helps the model place you.
5.2 Align profile ↔ post
Make your headline, About, and Featured sections echo those pillars. Audit recent posts for drift; reduce off-topic filler that dilutes your signals.
5.3 Write for read time and replies
Ask for specific experiences or counter-points. Aim for comments that add detail – those correlate with dwell and quality.
5.4 Seed the first hour
Get the right people reading early (DM a small circle, notify collaborators). Early relevant interactions are still a strong quality cue.
5.5 Use entities and structure
Clear titles, bullets, and named entities (brands, roles, locations). Avoid vague “thoughts on leadership” posts; be concrete.
5.6 Keep posting cadence steady
Regular signals help any recommender learn your audience; you don’t need daily posts – you need consistent ones your audience finishes and discusses.
6. What not to overthink
Hashtag stuffing: A couple of highly relevant tags help; long strings don’t.
Post length obsession: Short or long can work – what matters is clarity, finish rate, and meaningful responses.
“Engagement bait”: Empty “agree?” hooks may inflate comments short-term but weaken your expertise signal long-term.
7. Sources & further reading
360Brew research (pre-production paper): 150B decoder model; text interface; cross-task generalisation.
arXiv
LiRank (production ranking framework): ACM paper on LinkedIn’s industrial ranking models prior to/alongside LLM unification.
dl.acm.org
Dwell time & two-pass feed: LinkedIn engineering blog explains candidate generation and ranking with dwell signals.
LinkedIn
Feed relevance system overview (engineering): first-pass rankers and feed sources.
engineering.linkedin.com
Practitioner explainers/reactions: Hootsuite 2025 update; marketer/analyst round-ups discussing the LLM shift.
Social Media Dashboard
