跳到主要内容Skip to main content
投机取巧 Speculative Speculation

深度求索 DeepSeekDeepSeek　·　当前代current generation V4.1-FlashV4.1-Flash（2026-09）

DeepSeekDeepSeek模型家族谱系卡model family lineage card

MIT 许可 + KV cache 压缩路线鲜明；量化生态活跃（Redis 作者、llama.cpp 生态参与者 antirez 维护过其量化版，本站系列页有记录）。适合追新、能接受大模型的部署者。MIT-licensed with a distinctive KV-cache-compression roadmap; an active quantisation ecosystem (antirez, author of Redis and a llama.cpp ecosystem participant, maintains a quantised build; recorded on the series page). For deployers who track the frontier and can host large models.

本站已收录该家族的量化对比：This site covers this family's quantisations: DeepSeek V4DeepSeek V4

先说清楚：这个站的数据怎么来的First, where this site's data comes from

✅ 实测字段Measured fields

直接从 Hugging Face 公开 API 读取的客观数据：文件体积（字节）、上下文长度、架构、许可、下载量、更新时间。Objective data read straight from the Hugging Face public API: file sizes in bytes, context length, architecture, licence, downloads, last updated.

✅ 推导Derived

由实测字段按公开规则算出：每权重比特数（依 llama.cpp 命名标准）、档位归属、显存需求区间。推导规则在Computed from the measured fields by published rules: bits per weight (per the llama.cpp naming convention), tier assignment, VRAM requirement range. The rules are fully published in METHODOLOGY.md 完整公开，任何人可复算。 — anyone can recompute them.

❌ 我们不说的What we do not claim

哪个量化质量更高、哪个更快、哪个模型更强——这些需要实测，本站不跑 benchmark，所以不装作知道。凡我们给出的主观判断，一律显式标注「我们的口径」。Which quantisation is higher quality, which is faster, which model is stronger — these need measurement. This site does not run benchmarks, so it does not pretend to know. Every subjective judgement we do offer is explicitly marked as our own reading.

本站This site does 不下载、不托管任何模型权重文件not download or host any model weight files。所有数字都来自公开元数据，点击即可回到发布页核对。. Every figure comes from public metadata, and clicking it takes you back to the release page to verify. 采集时间Collected: 2026-10-30（数据快照，非实时） (a snapshot, not live)

中文生态China availability 国产 / 国内可直接用 / 备案状态 · 快照domestic / direct access / CAC filing · snapshot 2026-10-20

  • 出品属地Origin　国产domestic　出品方为深度求索（北京）。Made by DeepSeek (Beijing).
  • 国内下载Direct access　国内可直连direct OK　ModelScope 有 deepseek-ai 官方组织镜像（2026-10-20 实测 200）；Hugging Face 直连不可达。Official deepseek-ai org mirror on ModelScope (measured 200); Hugging Face direct unreachable.　·　国内镜像入口domestic mirror
  • 备案状态CAC filing　已备案Filed：Beijing-DeepseekChat-202404280016　公示名称listed as「求索对话DeepSeek Chat」　·　备案时间filed 2024-05-13　·　2024-08 批次第 152 行。Row 152 of the 2024-08 batch.

备案信息以网信办《生成式人工智能服务已备案信息》公示名单为准（按批次滚动更新，本站只抄编号不解读）；「国内可直连」是本站从大陆网络直连实测的结果，不是厂商承诺。Filing status is quoted from the CAC public registry (updated batch by batch; this site copies the number and does not interpret it). "Direct access" is this site's own reachability measurement from a mainland network, not a vendor promise.

代次Generations

  • V4.1-FlashV4.1-Flash2026-09当前代current

    当前代旗舰 Flash 线。官方定位「Pushing the Limits of KV Cache Compression」，MIT 许可。Current Flash-line flagship. Positioned as "Pushing the Limits of KV Cache Compression"; MIT-licensed.

    官方页面 ↗official page ↗
  • V4-Flash / V4-ProV4-Flash / V4-Pro2026-07

    V4 代两个分支：Flash（0731）与 Pro（0813）；本站收录的 deepseek-v4 系列在 V4.1-Flash 上。Two branches of the V4 generation: Flash (0731) and Pro (0813). This site's deepseek-v4 series tracks V4.1-Flash.

    官方页面 ↗official page ↗
  • V3.2V3.22025

    上一代主力（V3 系），MoE + 稀疏注意力路线从这一代开始成型。Previous mainline generation (V3 series); the MoE plus sparse-attention approach took shape here.

    官方页面 ↗official page ↗

尺寸档Size tiers

  • 单一MoE大杯：552B 骨干 + 196B Engram 条件记忆 + 视觉塔等（官方卡组件相加约 763B 总参，官方 README 未给单一总参数字），本地可跑性依赖量化与内存Single MoE tier: 552B backbone plus a 196B Engram conditional-memory table and a vision tower (the official card's components add up to roughly 763B total; the README gives no single total), local runability depends on quantisation and memory

许可Licences

官方能力声明（带出处）Official capability claims (with sources)

  • 官方 README：「a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens」——多模态 MoE、上下文上限 100 万 token。Official README: "a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens".出处 ↗source ↗
  • 官方 README：CED 架构下每 token 预填充激活 8B 参数、解码激活 16B（「8B parameters per token during prefill and 16B during decode」）。Official README: under the CED architecture, 8B parameters activate per token during prefill and 16B during decode.出处 ↗source ↗
  • 官方 README：DSpark 是投机解码的草稿生成（「DSpark speculative decoding (semi-autoregressive draft generation with confidence-scheduled verification)」），不是可单独使用的主权重。Official README: DSpark is speculative-decoding draft generation ("semi-autoregressive draft generation with confidence-scheduled verification"), not a standalone main weight set.出处 ↗source ↗

以上均为厂商官方页面原文或官方自评基准的转述；涉及横评结论的，本站不背书——本站不跑 benchmark。The above are quotations of official pages or transcriptions of vendor-run benchmarks; this site does not endorse cross-model conclusions — it runs no benchmarks.

出处Sources

核验快照：2026-10-30 · 本站不实测、不排名，未知就写未知。Verification snapshot: 2026-10-30 · no benchmarks, no rankings — unknown stays unknown.