DeepSeek-V4.1-Flash

The first model in DeepSeek's V4.1 architecture family and, per DeepSeek, the smallest of it: a natively multimodal 552B-parameter MoE that succeeds V4-Flash and, on DeepSeek's own tests, beats the 1.6T V4-Pro on performance, cost, and speed. The headline change is a Causal Encoder-Decoder (CED) layout: a 20-layer causal encoder feeds a 20-layer decoder whose global KV cache is projected from the encoder's final hidden states, so the model activates only 8B parameters per token during prefill and 16B during decode. Compressed Sparse Attention 2 (CSA2) shares main KV and indexer keys across layers in Full, Reindex, and Reuse modes with a hierarchical sparse indexer, and the main KV cache is stored in FP4, bringing the global KV footprint to 890 bytes per token, about a quarter of V4-Flash and 437 times smaller than DeepSeek-V1; SWA Bounded Replay cuts the persistent SSD cache to roughly an eighth. Decode FLOPs grow only about a quarter from 4K to 1M context. It keeps and revises the V4-era extensions: Single-Pass mHC, a 196B Engram conditional-memory table, and DSpark speculative decoding. Each MoE layer has 384 routed experts plus one shared, 6 active per token; a from-scratch DeepSeek-ViT encoder with 2D-RoPE handles images.

Pretrained from scratch on 45T multimodal tokens, with sparse attention trained at 64K and context extended to 1M tokens at the 34T mark; post-training is the standard SFT, RL, and on-policy-distillation recipe with the changes confined to large-scale synthetic agent tasks and environments. Reasoning effort is a continuous 1 to 100 dial; the API exposes max, high, and low tiers at 100, 75, and 50. Self-reported at max effort (DeepSeek Harness): DeepSWE v1.1 74.2 (Claude Opus 5 74.0, GPT-5.6 Sol 73.0, V4-Flash 54.4), Terminal-Bench 2.1 90.6, Terminal-Bench 4.0 31.2 (Opus 5 51.8), GPQA Diamond 90.9, HLE 36.8, Codeforces 3471, MathArena Apex 65.6, CyberGym 88.1, Automation-Bench 54.8, and Agents' Last Exam 31.8. MIT license; served as deepseek-flash at $0.30 per million input tokens and $1.20 per million output at peak, half that off-peak, with a 1M context and 384K max output. V4-Flash and V4-Flash-Vision-Exp were retired the same day, and V4-Pro traffic routes to V4.1-Flash from September 14 until a V4.1-Pro ships. Artificial Analysis scores it 39 on Intelligence Index v4.3 (reasoning, max effort), above V4-Pro-0813 at 36 and V4-Flash-0731 at 34, one point behind Qwen3.8 and otherwise behind only GLM-5.3 (45), Kimi K3 (44), and GLM-5.3-Flash (42) among open-weights models, at 198 tokens/s and a 1.06 s time to first token.