DeepSeek-V3.2 — full network dataflow (official inference reference)
Built from
deepseek-ai/DeepSeek-V3.2 inference/model.py
(ground truth), config.json, and
arXiv:2512.02556
(Gemini 3.1 Pro extraction, adversarially verified). Blue = GEMM (precision noted),
teal = elementwise/norm, red = attention core, orange = caches, purple = routing/index ops.
Drag to orbit · scroll to zoom · press ? for keys.
Key dimensions — DeepSeek-V3.2
hidden
H = 7168 ·
61 layers (3 dense + 58 MoE) · vocab
129280
MLA:
128 heads · q_lora
1536 · kv_lora
512 · nope 128 + rope 64 · v 128
KV/token:
512 c_kv (FP8) + 64 k_pe — MQA-absorbed decode
indexer:
64 heads × 128 · Σ w·ReLU(q·k) in FP8 ·
top-2048
MoE:
256 experts, top-8 via
4/8 groups · inter 2048 · shared ×1 · ×2.5
FP8 e4m3 block 128×128 (ue8m0) · YaRN 163840 · MTP ×1 (paper; absent in ref code)
FC / GEMM (precision as labeled)
elementwise / norm / reshape
attention / score core
KV cache / persistent state
routing / gating / speculative ops
embeddings / logits / block refs
─ solid: dataflow · ┄ dashed: control/index · ┄ orange: cache update