DeepSeek V4.1 architecture halves long-input prefill compute with novel causal encoder-decoder design aligned for long-running agent systems
From the report: DeepSeek V4.1 separates its 40-layer model into a 20-layer causal encoder and 20-layer decoder, both causal. Prompt tokens only go through the encoder, while generated tokens go through both, resulting in 8B active params for prefill vs. 16B for decode. The model uses a global KV cache shared and reused across layers via three modes: Full, Reindex, and Reuse, drastically…
0 pts