摘要

DeepSeek 发布 V4.1-Flash——新架构家族中最小的模型,原生多模态视觉:552B backbone(safetensors 总参数 763B,含 196B 的 Engram 条件记忆模块),每 token 激活 prefill 8B、decode 16B,1 共享 + 384 路由专家、激活 6 个。CED 设计把 40 层组织为 20 层因果 encoder + 20 层 decoder,decoder 的 KV cache 由 encoder 最终隐状态投影而来;CSA2 提供三种静态注意力模式(Full/Reindex/Reuse)与层级稀疏索引器,主 KV 采用 FP4(E2M1)缓存、每 token 890 字节(约为 V4-Flash 的 1/4;SWA bounded replay 把常驻 KV 再压到约 1/8),另有 Single-Pass mHC kernel 与 DSpark 推测解码。预训练 45T 多模态 token;上下文扩展至 1M。基准(reasoning_effort=100):Terminal-Bench 2.1 90.6、TB 3.0 30.0、TB 4.0 31.2、DeepSWE v1.1 74.2(V4-Flash 为 54.4)、GPQA Diamond 90.9、Codeforces 3471、HLE-with-tools 63.9、CyberGym 88.1、ExploitGym 15.3。MIT 许可、无门禁,day-0 支持 vLLM/SGLang/Docker Model Runner/Transformers(deepseek-recipe 库;无 Jinja chat template)。API:模型名改为 deepseek-flash(输入 $0.30 高峰 / $0.15 非高峰、输出 $1.20/$0.60 每百万 token;缓存命中 $0.006 高峰);V4-Flash 与 V4-Flash-Vision-Exp 从 API 退役,旧名暂路由至 V4.1-Flash;V4 Pro 在 2026-09-14 之后继续服务、计费不变;reasoning effort 为 1-100 的整数可控。约 1 天 75,774 下载 + 1,741 likes;采样时居 HF trending 第 1。

为什么重要
这是 DeepSeek 对 GLM-5.3-Flash 一代的同档回应,且以 Terminal-Bench 2.1(90.6 vs GLM-5.3 的 88.2)拿下开源权重 agentic coding 的最高分。架构与分数同样重要:encoder-decoder KV 投影 + FP4 KV + 条件记忆,从 serving 成本侧切入(混合线性注意力从另一侧切入),$0.30/$1.20 的 API 价格重画了廉价档。要盯 serving 摩擦:deepseek-recipe 运行时依赖对现有技术栈是新增的集成成本。
技术细节
发布 HF 仓库 2026-09-10T02:17:58Z 创建;MIT;image-text-to-text;约 1 天 75,774 下载 + 1,741 likes(2026-09-11T18:34Z 采样)
架构 552B backbone / safetensors 总参 763B(含 196B Engram 条件记忆);prefill 激活 8B / decode 激活 16B;MoE 1 共享 + 384 路由、激活 6;CED = 20 层因果 encoder + 20 层 decoder(KV 由 encoder 最终隐状态投影);CSA2 静态注意力 Full/Reindex/Reuse + 层级稀疏索引器;FP4 E2M1 主 KV,890 B/token(约 V4-Flash 的 1/4;SWA bounded replay 下约 1/8);Single-Pass mHC;DSpark 推测解码;DeepSeek-ViT 视觉(2D-RoPE、3x3 pixel-unshuffle)
训练 45T 多模态 token 预训练;64K 处稀疏注意力;34T 处扩展到 1M;SFT → RL → on-policy distillation;reasoning_effort 为 1-100 整数
基准 TB 2.1 90.6 / TB 3.0 30.0 / TB 4.0 31.2;DeepSWE v1.1 74.2(V4-Flash 54.4);GPQA Diamond 90.9;Codeforces 3471;HLE 36.8(纯文本 39.1);HLE w/tools 63.9;MathArena Apex 65.6;CyberGym 88.1;SEC-Bench Pro 62.8;ExploitGym 15.3;MMMU-Pro 56.5;DocVQA 95.6
API 模型名 deepseek-flash;输入 $0.30 高峰 / $0.15 非高峰、输出 $1.20/$0.60 每百万;缓存命中 $0.006 高峰;并发 2500;最大输出 384K;V4-Flash 与 V4-Flash-Vision-Exp 退役(旧名暂路由);V4 Pro 2026-09-14 后继续、计费不变
Serving day-0 vLLM / SGLang / Docker Model Runner / Transformers;deepseek-recipe Rust 库 + Python 参考实现;无 Jinja chat template
后续更新
2026-09-13 2026-09-13 run, serving-ecosystem check: vLLM now carries 63 issues/PRs mentioning V4.1-Flash (13 on 2026-09-12 alone - encoder CUDA-graph support #56625 17:54Z, mHC post-block folding into the delayed pre-projection #56633, ROCm gfx950 re-enable #56638; ROCm performance RFC #56506 opened 9/11); SGLang shows 32 matches (MI350X/gfx950 support #39186, Mooncake Engram KV backend with CUDA Graph #39205, prefill context parallelism #39179) plus a third-party report that fp8 wo_a absorb GEMM output is silently ~25% wrong when DEEPGEMM_SCALE_UE8M0=false on SM121 (#39193, 06:21Z). The HF model card still ships no serving recipe (community request in #13); deepseek-ai/deepseek-recipe is a prompt-format conversion library, not serving. 15+ third-party deployment repos created/updated on 9/12 target DGX Spark/GB10, Mac Studio 512GB and EXL3 quantizations; LiteLLM pricing PRs #40919/#40920 (9/12 23:36Z) add V4.1-Flash across five providers. Downloads @2026-09-13T00:03Z: 140,636 (+85.6% vs ~75.8k @9/11, ~2.5 days after release) + 2,005 likes. No third-party Terminal-Bench/SWE run yet. Stays TRIAL.
2026-09-14 2026-09-14 report run: vLLM main merged DeepSelect sparse-indexer TopK (#56464; 191 us vs 616 us on GB200 at batch 256 / 1M KV, kernel-only) and production paths for Engram DP sharding, async CPU-offload prefetch and shared-memory tables (#56512; two-node TP1xDP8 validation, <=0.9% reported TTFT gain). Not yet tagged. Recommendation stays TRIAL.
标签
foundation-modelopen-weightsdeepseekmultimodal1m-contextkv-cachemoeapi-pricing