DeepSeek released DeepSeek-V4.1-Flash, a 552-billion-parameter multimodal Mixture-of-Experts model with a Causal Encoder–Decoder layout and a one-million-token context window, positioned as the smallest member of a new architecture family and available via API as deepseek-flash. The 40-layer Transformer splits into a 20-layer causal encoder and a 20-layer decoder, activating about 8 billion parameters per token during prefill and about 16 billion during decode. Compressed Sparse Attention 2 assigns Full, Reindex, or Reuse modes across layers, and FP4 main KV caching shrinks the global KV footprint to roughly 890 bytes per token—about one-quarter of DeepSeek-V4-Flash—while SWA Bounded Replay cuts persistent SSD KV demand to about one-eighth. Training covered about 45 trillion multimodal tokens, with sparse attention at 64K sequences and context extended to 1M later in pretraining, followed by supervised fine-tuning, reinforcement learning, and on-policy distillation. On agentic instruct benchmarks at maximum reasoning effort, DeepSeek reports Terminal-Bench 2.1 at 90.6, DeepSWE v1.1 at 74.2, and AutomationBench at 54.8, alongside native vision via a from-scratch DeepSeek-ViT encoder. Weights and inference recipes are published on Hugging Face under an MIT license