Open Models Almanac
Methodsmodel architecturetoken mixersparse attentionsparse attention indexer

specific method · filed under model architecture

IndexShare

IndexShare is a sparse-attention mechanism that computes which earlier tokens each layer attends to once and reuses those indices across layers.1 Instead of each attention layer independently selecting tokens to focus on, it shares the indices among layers, cutting the cost of long-context attention.2

Trade-off It reduces per-token computational cost at long context while maintaining strong accuracy.3

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. computing sparse attention indices once and reusing them across every four attention layers — GLM-5.2 Explained: Z.ai Open-Weights Frontier Model - TechNow
  2. every attention layer must independently compute which earlier tokens to focus on, an expensive operation — GLM-5.2 Explained: Z.ai Open-Weights Frontier Model - TechNow
  3. this reduces per-token computational cost by 2.9× at 1M-token context while maintaining strong accuracy — GLM-5.2 Explained: Z.ai Open-Weights Frontier Model - TechNow
sources
3
model
1
lab adopt it
1
strongest
core

Documented in

Z.ai (Zhipu AI)

Evidence

3 quotes from 3 sources, one card per model, each opening with its strongest.

How sources treat it
used 1core 2

GLM-5.2

Z.ai (Zhipu AI)adopted3 sources

We propose IndexShare, which reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9× at a 1M context length.

coremodel architecturein GLM-5.2Z.ai
2 more from this model

We propose IndexShare, which reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9× at a 1M context length.

usedmodel architecturein GLM-5.2Z.ai

the most technically significant innovation in GLM-5.2 is a mechanism Z.ai calls IndexShare

coremodel architecturein GLM-5.2Z.ai

Further reading

Picked by hand, not extracted: where to read more, not evidence for anything on this page.

↑