DeepSeek-V4.1-Flash: 552B backbone MoE, 1M token context. Causal Encoder-Decoder (CED): 20-layer encoder + 20-layer decoder; 8B active prefill, 16B decode.
CSA2 (Compressed Sparse Attention 2): Full/Reindex/Reuse modes share KV and sparse-attention indices across layers. FP4 KV caching (E2M1, E4M3 scale per 16 channels) reduces global KV to 890 bytes/token — ~1/4 of V4-Flash.
SWA Bounded Replay: persistent KV on SSD ~1/8 of V4-Flash by replaying recent n_win tokens.
Pretrained on 45T multimodal tokens. Agent benchmarks include Terminal-Bench 2.1 90.6, DeepSWE v1.1 74.2, CyberGym 88.1 (harness-dependent).