Definition

Transformer architecture splitting 40 layers into 20-layer causal encoder + 20-layer decoder. Decoder global KV cache projected from final encoder hidden states rather than per-layer derivation.

Key Points

  • Used in deepseek-v41-flash: 8B parameters active during prefill, 16B during decode
  • Improves cost efficiency for input-heavy agentic-ai workloads
  • Enables smaller active compute during long-context prefill phase

Sources