DeepSeek released V4.1-Flash: a 552B causal encoder-decoder that activates 8B params on prefill and 16B on decode

DeepSeek (Hugging Face model card)·high signalXBlueskyLinkedInCopy link

DeepSeek pushed V4.1-Flash to Hugging Face at 02:17 UTC on 2026-09-10, an MIT-licensed multimodal MoE with 552B backbone params and a 1M-token context. The Causal Encoder-Decoder architecture projects the decoder's global KV cache from final encoder hidden states rather than per-layer, so only 8B params activate during prefill and 16B during decode; SWA Bounded Replay cuts the persistent KV cache footprint to roughly 1/8 of V4-Flash. It posts Terminal-Bench 2.1 at 90.6, DeepSWE v1.1 at 74.2 and Codeforces 3471, and the card includes a scaffold comparison showing its own minimal harness (90.6) beating Claude Code (88.0) and Codex (84.1) on the same model.