/Tech15h ago

GLM-5.2 launches as a 744B open-weight MoE model with a 1-million-token context window

It uses an IndexShare mechanism to reduce inference costs.

3897812448942.1K

#182

Original post

Sebastian Raschka@rasbt#182inTech

Just caught up with the recent GLM-5.2 release. The best open-weight model today.

Architecture-wise, it's build on the GLM-5 and GLM-5.1 architecture that I covered previously, which means it's reusing the Multi-head Latent Attention (MLA) and DeepSeek Sparse Attention (DSA) mechanisms from DeepSeek V3.2. (I wrote about it here: https://magazine.sebastianraschka.com/p/technical-deepseek)

What's new is that they added an IndexShare mechanism. (That's a cross-layer reuse trick for DSA where instead of recomputing the sparse-attention top-k indexer in every layer, GLM-5.2 runs the full indexer only once every four layers and lets the following layers reuse those selected token indices. This keeps the same DSA idea but makes 1M-token inference much cheaper.)

7:15 AM · Jun 18, 2026 · 42.5K Views

Sentiment

Many users praise GLM-5.2's IndexShare efficiency as a smart optimization that cuts 1M-token inference costs while making it the leading open-weight model.

Pos

100.0%

Neg

0.0%

14 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

SEBASTIANRASCHKA.COMVia

#182

Posts from X

Most Activity

VIEWS1.2KREPLIES3

Kai@Kai045l

@rasbt Isn't Kimi K2.7 the best open-weight model today?

14h1.2K2

BOOKMARKS4

Ethan@ethankongee

Let me add some more context on Sparse Attention and IndexShare.

In a classic attention block, a token has to attend to all previous tokens. But this becomes computationally expensive as the context window gets larger, so it’s not very scalable. GLM-5.1 has 200K token context window and GLM-5.2 has 1M. They need to do something to fix it or the performance would degrade.

What if we tell the model to only attend to the important tokens? That’s the core idea behind sparse attention.

Instead of looking at the entire context every time, sparse attention first selects a smaller shortlist of relevant tokens, and the model only attends to that shortlist.

But computing that shortlist is also quite expensive. If every layer of the model has to run its own indexer, it still slows down the model. It turns out that nearby layers in a model often care about similar tokens, so http://Z.ai decided to share the same index across every 4 layers. This is IndexShare mechanism. The first layer in a stack of 4 layers runs the indexer and computes the top-k indices, while the other 3 layers reuse that same index.

13h46734

LIKES15

Sebastian Raschka@rasbt

@Kai045l that was last week :D

12h67015

RETWEETS112

Sebastian Raschka@rasbt

Just caught up with the recent GLM-5.2 release. The best open-weight model today.

15h42.5K982489

Sebastian Raschka@rasbt

@vibecoder_dc I don’t disagree. Let’s say most capable open-weight LLM when averaged over all major benchmarks (reasoning, coding, logic, tool use, math, knowledge)

11h5049

Tim Riggins@MrTimRiggins

@rasbt ah man, i need to get up to speed, haven't tried any GLM models yet 😅

15h1.1K4

DC｜use.fo@vibecoder_dc

@rasbt 'Best open-weight' is doing a benchmark claim without saying 'on what.' Like saying 'best car' without saying 'for what.' Which benchmark, which hardware, which use case?The 'best open-weight' hides the benchmark. Like calling a car 'best' without 'for what.'

11h615

Jordi Mon Companys@JordiMonPMM

@rasbt I need your attention on Laguna, immediately!

12h141

Prince does AI@princedoesai

@rasbt ok that indexshare trick is actually smart

14h5851

Artale@artale93

@rasbt indexshare is such a smart optimization trick for 1m token inference, efficiency gains like this are exactly what open weight models need.

15h4431

Steven Cheng@stevencheng

@rasbt Reusing the indexer every four layers is just standard caching, my robot dogs do the same when tired, it is just efficient laziness.

12h2181

Ferbin@Ferbin08

@rasbt how does inference latency compare? sparse attention saves memory but the routing overhead usually kills speed gains at runtime

14h596

Alpha Batcher@alphabatcher

@rasbt in my honest opinion :

GLM-5.2 probably without any competition right now

11h273

Maryam@Sci_Tech_Eng

@rasbt Strong & pioneer...

13h243

Maryam@Sci_Tech_Eng

@rasbt Helpful ! Thanks Sebastian! 🙏

13h211

The Whizz AI@TheWhizzAI

@rasbt smart and too valuable

12h173

Mustafa Ergisi@mustafaergisi

@rasbt MLA plus DSA is becoming the default stack for the strongest open weights. The architecture race for serving efficiency has its current frontrunners.

10h541

Rasmus Schultz@mindplaydk

@rasbt while others just train a bigger model and pretend they're making progress, it seems GLM is actually improving the model 😊

8h136

踏空 Sidelined Capital@sidelined_cap

@rasbt Interesting that GLM-5.2 is borrowing from both MLA and sparse attention ideas. Open-weight progress now feels like architecture diffusion is moving almost as fast as model releases.

10h121

Akumu@Akumunokokoro

@rasbt Does the verbosity negate the inference gains? @ArtificialAnlys shows GLM-5.1 at 120M tokens and GLM-5.2 at 140M.

11h113