Tech report · 2026-09-10 · interactive model

DeepSeek-V4.1-Flash

552B MoE backbone, 40-layer Causal Encoder–Decoder, 890 bytes of global KV per token. Click any part of the stack — or any label — to read what it does and what it buys.

SELF-DESTRUCT IN 60:00
this exhibit is temporary · one hour only
drag · scroll · click a part · ← → to step · esc to close

Exhibit closed

This model was provisioned for one hour and has now been decommissioned. The container is gone; the architecture page in the wiki remains.

~/wiki/entities/deepseek-v41-flash-tech-report.md
assembling 40 blocks