552B backbone + 196B Engram / 8B active during prefill / 16B active during decode (MoE)
1,048,576 tokens
MIT
Official open-weight multimodal model released immediately after a two-day first-party API beta. DeepSeek-V4.1-Flash is trained from scratch rather than being another post-training update to V4-Flash. It uses a 40-layer Causal Encoder-Decoder architecture comprising a 20-layer causal encoder and 20-layer decoder, with 552B backbone parameters and 196B additional Engram conditional-memory parameters. The model activates 8B parameters per token during prefill and 16B during decoding, using one shared expert and 384 routed experts per MoE layer with six routed experts selected per token. Compressed Sparse Attention 2 introduces Full, Reindex, and Reuse attention modes, cross-layer KV and index reuse, and hierarchical sparse indexing. Combined with FP4 main-KV caching, this reduces the always-resident global KV cache to 890 bytes per token, approximately one quarter of DeepSeek-V4-Flash, while SWA Bounded Replay reduces the persistent SSD or host-memory KV footprint to approximately one eighth. Additional components include Single-Pass mHC residual mixing, two sparsely accessed Engram modules, and integrated DSpark speculative decoding. A 32-layer DeepSeek-ViT vision encoder with 2D RoPE, 14-pixel patches, 3x3 pixel-unshuffle downsampling, and up to 1,024 image tokens provides native image understanding. Pretraining used a 45T-token multimodal corpus, with sparse attention trained from scratch at 64K sequence length and context extended to 1,048,576 tokens after 34T tokens. Post-training follows SFT, reinforcement learning, and on-policy distillation while scaling automatically synthesized agent tasks, environments, and rollouts; reasoning effort is continuously controllable from 1 to 100. At maximum effort, DeepSeek reports 90.6 on Terminal-Bench 2.1, 30.0 on Terminal-Bench 3.0, 31.2 on Terminal-Bench 4.0, 74.2 on DeepSWE v1.1 with mini-SWE-agent, 72.6 with DeepSeek Harness Minimal, 88.1 on CyberGym, 54.8 on AutomationBench, and 31.8 on Agent's Last Exam. The checkpoint is published across 48 mixed FP4/FP8 safetensors shards totaling approximately 510 GB. It does not include a Jinja chat template; DeepSeek instead provides reference conversation encoding, maintained deepseek-recipe libraries, local-inference guidance, and reproducible DeepSWE evaluation instructions.
https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/DeepSeek_V41_Tech_Report.pdf