Model architecture atlas · September 2026

DeepSeek V4.1 Flash

A clickable map of the multimodal 40-layer Causal Encoder-Decoder, CSA2 cross-layer cache sharing, hierarchical sparse indexing, Single-Pass mHC, Engram conditional memory, multimodal MoE routing, and DSpark drafting.

552B + 196Bbackbone + Engram parameters
8B / 16Bactive parameters · prefill / decode
20 + 20causal encoder + decoder layers
1Mmaximum context length
890 B/tokenreported global KV footprint

System flow

Images and text converge before the 20-layer encoder and 20-layer decoder.

Click a shape for details · drag to pan · wheel to zoom

System symbols

Primary sources

Shapes and schedules are taken from the official model config, technical report, and reference inference implementation. The diagrams show logical model structure rather than deployment-specific storage layouts.

Technical report

CED, CSA2 modes, hierarchical indexing, Single-Pass mHC, Engram, DSpark, and KV formats.

Open report ↗ Official config

Exact dimensions, layer source lists, compression ratios, experts, vision tower, and draft settings.

Open config ↗ Reference model

Executable compressor, indexer, attention, Engram, mHC, MoE, vision merge, and DSpark dataflow.

Open model.py ↗