DeepSeek-V4.1-Flash is designed around KV bytes rather than FLOPs. Its 40 layers (d = 5,120) form a causal encoder-decoder: the 20 decoder layers take their global KV from the encoder's final hidden state instead of their own, so prefill mostly runs the encoder (8B active parameters per token in prefill, 16B in decode). Attention is CSA2, which shares one compressed KV cache and its indexer keys across groups of layers (Full, Reindex and Reuse modes). Only 4 of 40 layers write global KV, three encoder layers at 2× compression and one uncompressed decoder layer, and a hierarchical indexer bounds the decoder's re-scoring to a 16,384-position candidate pool. V4's HCA layers are gone. With the main KV cached in FP4, global KV is 890 bytes per token, about a quarter of V4-Flash's.
The rest of V4 is kept and simplified: shared-KV MQA with a 128-token window branch and sink logits, √softplus routing over 384 experts (6 active plus 1 shared, no hash-routed layers), and Single-Pass mHC, which reuses the previous sublayer's mixing coefficients so one fused kernel reads the residual once. New are two Engram n-gram memory modules (196B parameters on top of the 552B backbone), separate routing biases for image tokens, and the DSpark drafter, trained after pretraining in place of MTP.
The model is natively multimodal: a from-scratch DeepSeek-ViT feeds images into the token stream from the start of 45T-token pretraining, which ran sparse attention from scratch at 64K with no dense warm-up. DeepSeek released it on 2026-09-10; a post-training update, DeepSeek-V4.1-Flash-1017, followed on 2026-10-17 with no architecture change described.