Model techniques map
Techniquesmodel architecturetoken mixerlinear attention & state spacegated delta network

specific method · filed under model architecture

Gated DeltaNet

A linear-attention method using a gated recurrence and delta-rule state updates, with recurrent state used to represent sequence history.

Also called Gated Delta Net (GDN) module, Gated Delta Network linear attention, linear attention via Gated Delta Networks, gated delta networks hybrid attention, Gated DeltaNet mechanism, Gated DeltaNet (GDN).

sources
13
models
5
labs adopt it
2
strongest
core

How sources treat it

One count per evidence span, weakest treatment to strongest.

evaluated 1used 2core 10

Documented in

Further reading

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

In the code

Checked as “Gated DeltaNet layers” against each model’s own modeling code and config.json, independently of its documents. A model’s page shows the quoted lines.

ModelCodeIts documentsWhat the code does
GLM-5.3-Flash in its codenot statedGlm5NextTextLinearAttention implements Kimi Delta Attention (a gated delta-rule variant) with a forget gate decay and delta-rule state update, used in linear_attention layers. modeling_glm5_next.py#L508
DeepSeek-V4.1-Flash not in its codenot statedNo gated delta-rule recurrent mixer code exists in the given files; attention layers are exclusively sparse/sliding-window MLA.
Hy4-preview not in its codenot statedNo gated delta-rule recurrent mixer code or config keys (linear_num_key_heads, linear_conv_kernel_dim) exist in this model.
DeepSeek-V4-Flash-0731 not in its codenot statedNo gated delta-rule recurrent module exists in the code; all layers use DeepseekV4Attention softmax attention.
NVIDIA-Nemotron-3-Ultra-550B-A55B not in its codenot statedThe recurrent layers implement the Mamba-2 selective SSM (chunked scan with segment sums), not a gated delta-rule mixer like Gated DeltaNet or Kimi Delta Attention. modeling_nemotron_h.py#L366
MiMo-V2.6-Flash not in its codenot statedNo gated delta-rule recurrent mixer is implemented; only softmax attention layers exist.
DeepSeek-V4-Flash not in its codenot statedNo gated delta-rule recurrent module exists in the code; all layers use DeepseekV4Attention (softmax attention with optional compression).
MiMo-V2.5 not in its codenot statedNo gated delta rule / recurrent state update mechanism is implemented in the code.
GLM-5.3 not in its codenot statedNo gated delta-rule recurrent mixer or linear_num_key_heads/linear_conv_kernel_dim config exists in this codebase.
Hy3 not in its codenot statedNo gated delta-rule recurrent module is implemented anywhere in the file
GLM-5.2 not in its code evaluatedNo delta-rule recurrent state or decay gate mechanism is implemented anywhere in the modeling file.
MiniMax-M3 not in its codenot statedNo gated delta-rule recurrent mixer is implemented; only standard attention and the lightning indexer exist.
DeepSeek-V3.2 not in its codenot statedNo gated delta rule / recurrent state mixer is implemented; only MLA attention layers exist.
DeepSeek-V4-Flash-Vision-Exp not in its codenot statedNo gated delta-rule recurrent mixer is implemented anywhere in this file.
DeepSeek-V4-Pro not in its codenot statedNo gated delta-rule recurrent mixer implementation is present in the given files
DeepSeek-V4-Pro-0813 not in its codenot statedNo gated delta-rule recurrent module exists in the code; layer_types only cover sliding/CSA/HCA softmax-attention variants.
Gemma 4 31B not in its codenot statedNo gated delta-rule recurrent mixer is implemented; the model uses only softmax attention layer types.
Inkling not in its codenot statedNo gated delta-rule recurrent mixer is implemented; the short convolution (sconv) is a plain depthwise causal conv1d with residual add, not a delta-rule/decay-gated state update.
Kimi K3 in its code usedKimiDeltaAttention implements Kimi Delta Attention (a Gated DeltaNet variant) with a delta-rule recurrent update, decay gate (A_log/dt_bias), and short convolutions on q/k/v, applied to the kda_layers. modeling_kimi_linear.py#L520
Laguna-S-2.1 not in its codenot statedNo delta-rule recurrent state update or decay gate code exists; all layers are standard softmax attention.
MiMo-V2.5-Pro not in its codenot statedNo gated delta-rule recurrent mixer implementation or config keys (linear_num_key_heads, linear_conv_kernel_dim) are present.
MiMo-V2.6-Pro not in its codenot statedNo gated delta-rule recurrent module is implemented; all token mixing is standard softmax attention.
NVIDIA-Nemotron-3.5-Lightning-30B-A3B not in its codenot statedThe linear-attention layers implement Mamba-2 style SSM scan (segment sum, chunked SSD), not a gated delta rule / Kimi Delta Attention recurrence.
Qwen3.5-397B-A17B in its code coreLinear-attention layers implement the gated delta rule (Qwen3_5MoeGatedDeltaNet) with a decay gate g and delta-rule chunked/recurrent update, configured via linear_num_key_heads/linear_conv_kernel_dim. modeling_qwen3_5_moe.py#L621
Qwen3.6-35B-A3B in its code coreQwen3_5MoeGatedDeltaNet implements the gated delta rule with a decay gate (A_log/dt_bias) and beta-weighted delta update, used for all linear_attention layers. modeling_qwen3_5_moe.py#L621
Qwen3.8-Flash-Next in its code coreQwen4ExpTextGatedDeltaNet implements the gated delta rule with a decay gate (A_log/dt_bias) and delta-rule state update, used on linear_attention layers. modeling_qwen4_exp.py#L581
Step-3.7-Flash not in its codenot statedNo gated delta-rule recurrent module exists in the given code; layer_types only distinguishes full and sliding softmax attention.
gpt-oss-120b not in its codenot statedNo gated delta-rule recurrent layers are implemented; all layers use GptOssAttention.

Evidence

13 spans quoted from the sources, strongest treatment first.

It combines Mamba2’s gated decay mechanism with a delta rule for updating hidden states.

coremodel architecturein Qwen3.5-397B-A17BAlibaba

Gated DeltaNet:- Number of Linear Attention Heads: 64 for V and 16 for QK

coremodel architecturein Qwen3.5-397B-A17BQwen

Gated DeltaNet:- Number of Linear Attention Heads: 32 for V and 16 for QK

coremodel architecturein Qwen3.5-35B-A3BQwen

The 3:1 interleave of Gated DeltaNet to full-attention layers is the same ratio Qwen3.5 uses, keeping most layers on linear attention whose state does not grow with sequence length.

coremodel architecturein Qwen3.8-Flash-NextQwen

GDN compresses the prefix into a fixed-size recurrent state and updates that state according to the current content

coreunclearin Qwen3.8-Flash-NextQwen

It uses the same gated delta networks hybrid attention as its MoE siblings

coremodel architecturein Qwen3.6-27BQwen

three of every four layers use Gated DeltaNet to compress history

coremodel architecturein Qwen3.8-Flash-NextQwen

linear attention via Gated Delta Networks

coremodel architecturein Qwen3.5-397B-A17BAlibaba

Gated DeltaNet: - Number of Linear Attention Heads: 32 for V and 16 for QK

coremodel architecturein Qwen3.6-35B-A3BQwen

Gated DeltaNet (GDN) compresses the history efficiently

coremodel architecturein Qwen3.8-Flash-NextQwen

similar to the Gated DeltaNet used for Olmo Hybrid

usedmodel architecturein OLMo HybridAi2

this architecture includes the Gated Delta Net (GDN) module, which is particularly effective for accelerating the modeling of long audio-video sequences.

usedmodel architecturein Qwen3.5-OmniQwen

Gated DeltaNet (GDN) [54]: A linear attention variant that replaces the quadratic softmax attention computation with a gated linear recurrence, reducing the computational cost of attention from quadratic to linear in sequence length.

evaluatedmodel architecturein GLM-5Z.ai

Filed alongside

Other methods under model architecture :: token mixer :: linear attention & state space :: gated delta network.