552B backbone + 196B Engram / 8B active during prefill / 16B active during decode (MoE)
1,048,576 tokens
MIT
DeepSeek describes DeepSeek-V4.1-Flash-1017 as a minor update to the September DeepSeek-V4.1-Flash checkpoint that delivers substantial improvements across reasoning, coding, agentic, and visual benchmarks. The company says the update also reduces task execution time and output length, allowing users to complete tasks faster and at lower cost while advancing the performance-cost Pareto frontier across multiple agent harnesses. Publisher-reported results versus the September checkpoint include HLE without tools 42.6 versus 36.8, Terminal-Bench 4.0 42.3 versus 31.2, DeepSWE v1.1 77.0 versus 74.2, ProgramBench 29.1 versus 20.3, NL2Repo-Bench 68.2 versus 65.4, Automation-Bench 60.5 versus 54.8, Agents' Last Exam 35.4 versus 31.8, and Apex Bench 72.1 versus 64.6. Tool-assisted Chartography, BabyVision, and ZeroBench-main rise to 83.0, 92.5, and 51.0 respectively. Terminal-Bench 4.0 uses the Claude Code harness, while DeepSWE uses Mini-SWE-V2. The update retains the V4.1 architecture and 1,048,576-token context window: a 40-layer Causal Encoder-Decoder with a 20-layer causal encoder and 20-layer decoder, 552B backbone parameters, 196B of Engram conditional-memory parameters, 8B activated during prefill, and 16B activated during decoding. Each MoE layer uses one shared expert and 384 routed experts with six selected per token. The architecture combines Compressed Sparse Attention 2, FP4 main-KV caching, SWA Bounded Replay, Single-Pass mHC residual mixing, two sparsely accessed Engram modules, and integrated DSpark speculative decoding. A 32-layer DeepSeek-ViT encoder provides native image understanding with up to 1,024 image tokens. DeepSeek does not disclose the update's additional training data, token budget, compute, or exact post-training recipe. Official Hugging Face and ModelScope repositories provide the MIT-licensed checkpoint in 48 mixed-precision safetensors shards totaling approximately 510.3 GB. The release does not include a Jinja chat template; DeepSeek instead supplies reference encoding code and the maintained deepseek-recipe toolkit for multimodal conversations, thinking, numeric reasoning effort, tool calls, and Chat Completions or Responses API conversion. Recommended settings are temperature 1.0, top-p 1.0, a 1M-token context window, and at least 256K maximum output tokens.
https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/DeepSeek_V41_Tech_Report.pdf