Architecture atlas
Model Architectures
Interactive architecture references with clickable components and tensor contracts.
Causal encoder-decoder
DeepSeek V4.1 Flash
Trace DeepSeek V4.1 through its multimodal causal encoder-decoder, CSA2 operating modes, hierarchical sparse indexing, Single-Pass mHC, Engram memory, MoE routing, and DSpark drafting.
- CSA2
- Sparse indexer
- Single-Pass mHC
- DSpark
Sparse + linear attention
GLM 5.3
Compare GLM 5.3 Base and Flash through DSA, KPool, KDA recurrence, mHC residual mixing, sparse experts, NextN, and native vision.
- DSA / KPool
- KDA
- mHC
- Multimodal
Hybrid recurrent architecture
Kimi K3
Explore Kimi K3 from token flow down to MLA, short convolution, gated delta-rule state updates, MoE routing, and kernel tensor shapes.
- MLA
- Delta rule
- MoE
- Kernel internals
Million-token architecture
DeepSeek V4
Trace DeepSeek V4 through CSA and HCA, the lightning indexer, manifold-constrained residuals, DeepSeekMoE, and multi-token prediction.
- CSA / HCA
- mHC
- DeepSeekMoE
- MTP