Model techniques map
Techniquesmodel architecturenormalization & residual

specific method · filed under model architecture

Manifold-Constrained Hyper-Connections

A Hyper-Connections variant that constrains residual mapping to the manifold of doubly stochastic matrices.

Also called mHC.

sources
8
models
4
labs adopt it
2
strongest
core

How sources treat it

One count per evidence span, weakest treatment to strongest.

evaluated 1used 4core 3

Documented in

Further reading

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

Evidence

8 spans quoted from the sources, strongest treatment first.

The model adopts Manifold-Constrained Hyper-Connections to improve scaling efficiency.

coremodel architecturein GLM-5.3-FlashZ.ai

The model also adopts Manifold-Constrained Hyper-Connections (mHC)

coremodel architecturein GLM-5.3-FlashZ.ai

DeepSeek-V4 series incorporate Manifold-Constrained Hyper-Connections (mHC) to strengthen the conventional residual connections

coremodel architecturein DeepSeek-V4DeepSeek

Manifold-Constrained Hyper-Connections (mHC): constrains residual mapping onto the manifold of doubly stochastic matrices (Birkhoff polytope), enhancing signal propagation stability while preserving model expressivity.

usedmodel architecturein DeepSeek-V4DeepSeek

mHC is confirmed in V4 as described in the December 2025 paper

usedmodel architecturein DeepSeek-V4DeepSeek

incorporate mHC to strengthen conventional residual connections

usedmodel architecturein DeepSeek-V4-FlashDeepSeek

Manifold-Constrained Hyper-Connections (mHC): We incorporate mHC to strengthen conventional residual connections, enhancing stability of signal propagation across layers while preserving model expressivity.

usedmodel architecturein DeepSeek-V4DeepSeek

mHC (Xie et al., 2025) uses a sigmoid and additionally constrains to a manifold of doubly stochastic matrices.

evaluatedunclearin Qwen3.8-Flash-NextQwen

Filed alongside

Other methods under model architecture :: normalization & residual.