AI News AI资讯 1d ago Updated 2h ago 更新于 2小时前 55

DeepSeek v4.1-Flash: 763B-P8B-D16B novel causal Encoder–Decoder architecture with vision marks the Return of the Whale DeepSeek v4.1-Flash:763B-P8B-D16B创新因果编码器-解码器架构融合视觉能力,巨鲸归来

DeepSeek launched V4.1-Flash as an open-weight flagship featuring a radically new causal Encoder-Decoder architecture, retiring V4 Pro entirely despite the incremental-sounding version name The model uses an asymmetric active-parameter design (8B for prefill/input, 16B for decode/output) within a 763B total-parameter MoE, achieving ~1-2% sparsity and a KV cache footprint up to 8x smaller than V4 Flash Despite being labeled "v4.1," independent benchmarks (Artificial Analysis Intelligence Index sc DeepSeek发布V4.1 Flash开源模型,采用因果编码-解码器架构,总参数763B但仅激活8B输入/16B输出 模型实现极致推理效率,KV cache占用仅为V4 Flash的1/8,支持1M token上下文 定价极具竞争力:$0.30/1M输入 tokens,$1.20/1M输出 tokens,缓存输入仅$0.006/1M 在Vals Index开源模型排行榜登顶,超越Kimi K3,Artificial Analysis评分达40分 这是DeepSeek自R1后重返SOTA研究的首个重要成果,标志开源模型竞争格局变化

82
Hot 热度
75
Quality 质量
78
Impact 影响力

Analysis 深度分析

TL;DR

  • DeepSeek launched V4.1-Flash as an open-weight flagship featuring a radically new causal Encoder-Decoder architecture, retiring V4 Pro entirely despite the incremental-sounding version name
  • The model uses an asymmetric active-parameter design (8B for prefill/input, 16B for decode/output) within a 763B total-parameter MoE, achieving ~1-2% sparsity and a KV cache footprint up to 8x smaller than V4 Flash
  • Despite being labeled "v4.1," independent benchmarks (Artificial Analysis Intelligence Index score of 40; #1 open-weight on Vals Index ahead of Kimi K3) confirm it is a generational leap, not a minor iteration
  • Pricing is dramatically competitive at $0.30/1M input tokens and $1.20/1M output tokens ($0.006/1M cached input), with an additional 50% off-peak discount, making it the cheapest model in the open-weight top 10
  • Key supporting innovations include Sliding-Window Attention Bounded Replay for memory efficiency, native text+image multimodal support in a single model, and a 1M-token context window, all under an MIT license

Why It Matters

DeepSeek's return to publishing SOTA research with a fundamentally new architecture signals a strategic pivot back toward open-weight leadership after ceding ground to GLM and Kimi, reshaping the competitive landscape for cost-efficient, long-context inference. The causal Encoder-Decoder design with prefill/decode separation represents a significant architectural departure from the all-Decoder paradigms dominant in recent large-language model development, potentially influencing future open-model trajectories. For practitioners, V4.1-Flash offers a rare combination of top-tier open-weight performance, extreme inference efficiency, and rock-bottom pricing that could redefine cost-performance trade-offs for production agent workloads and long-context applications.

Technical Details

  • Causal Encoder-Decoder Architecture: DeepSeek V4.1-Flash introduces a novel causal Encoder-Decoder structure that splits active computation between input (prefill) and output (decode) phases—8B active parameters for prefill and 16B active parameters for decode within a 763B total-parameter Mixture-of-Experts backbone, yielding 1-2% effective sparsity
  • KV Cache Optimization: Combined with Sliding-Window Attention Bounded Replay, the architecture reduces KV cache footprint to approximately 1/8th that of V4 Flash, dramatically lowering memory costs for long-running and multi-turn agent workloads
  • Multimodal & Context: Single model natively supports both text and image inputs (vision integrated without requiring a separate model), with a 1M-token context window and up to 384 max output tokens
  • Licensing & Availability: Released under MIT license with US/API availability via DeepSeek's first-party endpoint; rapidly adopted by infrastructure providers including Baseten (day-0 support) and Ollama (rolled out across Max, Team, and Pro tiers)
  • Post-Training Philosophy: DeepSeek's post-training approach aligns with Prof Jie Tang's perspective, emphasizing efficiency and creative context utilization over brute-force benchmark chasing

Industry Insight

  • DeepSeek's decision to title a generational architectural overhaul as "v4.1 Flash" is a deliberate signal that the most impactful advances may not always appear in headline版本号—practitioners who dismiss models based on version numbering alone risk missing genuine leaps, a point DeepSeek explicitly acknowledges
  • The prefill/decode asymmetric active-parameter design with near-zero KV cache overhead makes this model particularly well-suited for production agent systems that require long context retention across many turns, potentially lowering the barrier to deploying capable agents at scale
  • At $0.30/1M input tokens with cached input at $0.006/1M and off-peak discounts, V4.1-Flash compresses the cost-performance frontier so aggressively that it pressures all open-weight competitors to re-evaluate their pricing and architectural strategies, especially as more organizations shift toward open models for compliance and cost reasons

TL;DR

  • DeepSeek发布V4.1 Flash开源模型,采用因果编码-解码器架构,总参数763B但仅激活8B输入/16B输出
  • 模型实现极致推理效率,KV cache占用仅为V4 Flash的1/8,支持1M token上下文
  • 定价极具竞争力:$0.30/1M输入 tokens,$1.20/1M输出 tokens,缓存输入仅$0.006/1M
  • 在Vals Index开源模型排行榜登顶,超越Kimi K3,Artificial Analysis评分达40分
  • 这是DeepSeek自R1后重返SOTA研究的首个重要成果,标志开源模型竞争格局变化

为什么值得看

本文揭示了开源大模型向极致推理效率演进的关键技术路径,对开发者选择推理友好型模型具有直接参考价值。DeepSeek的架构创新为降低AI部署成本提供了新思路,值得AI从业者关注。

技术解析

架构创新:V4.1 Flash采用因果编码-解码器(Causal Encoder-Decoder)设计,实现prefill/decode分离——8B参数处理输入token,16B参数处理输出token,形成"763B-P8B-D16B"的稀疏MoE结构,稀疏度仅1-2%。

效率优化:结合Sliding-Window Attention Bounded Replay等技术,KV cache占用降至V4 Flash的1/8,大幅降低长上下文推理的显存和成本开销,特别适合长时间运行的Agent场景。

性能与成本:1M token上下文,支持文本+图像多模态输入,MIT开源协议。Artificial Analysis Intelligence Index评分40分,超过V4 Pro 0813版本,同时成本极低。

评测基准:Vals评测条件为1M上下文、384最大输出、温度1、默认top-p/top-k、高推理力度,该配置下成为最便宜的Top 10开源模型。

行业启示

开源模型差异化竞争:DeepSeek避开与传统闭源模型的正面参数竞赛,转向推理效率和成本控制,为开源模型找到了一条可持续的商业化路径。

架构创新驱动降本:prefill/decode分离和极端稀疏激活的技术路线证明,模型性能提升不一定依赖全参数增长,架构创新同样是重要的成本优化杠杆。

评测标准需要更新:当前benchmark难以完全捕捉V4.1 Flash在长上下文高效利用方面的优势,行业需要更细粒度的评测体系来衡量实际部署价值而非单纯的性能峰值。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 LLM 大模型 Multimodal 多模态 Product Launch 产品发布 Training 训练