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

月之暗面（Moonshot AI）Moonshot AI　·　当前代current generation Kimi K3Kimi K3（2026-06）

KimiKimi模型家族谱系卡model family lineage card

国产开放权重旗舰：K3 是官方口径的首个开放 3T 级模型，许可是自定义 Kimi K3 License。对话产品「Kimi 智能助手」在网信办首批备案（Beijing-MoonShot-20231016）；模型权重本身与备案是两回事，卡上分开写。The domestic open-weight flagship: K3 is, per the vendor, the first open 3T-class model, under a custom Kimi K3 License. The consumer chat product was filed in the CAC's first batch (Beijing-MoonShot-20231016); model weights and service filing are different things — the card keeps them apart.

本站已收录该家族的量化对比：This site covers this family's quantisations: Kimi K3Kimi K3

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　出品方为月之暗面（北京）。Made by Moonshot AI (Beijing).
  • 国内下载Direct access　国内可直连direct OK　ModelScope 官方镜像直连可达（2026-10-20 实测 200）；Hugging Face 直连不可达。Official ModelScope mirror reachable (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-MoonShot-20231016　公示名称listed as「Moonshot」　·　备案时间filed 2023-11-03　·　首批第 16 行（月之暗面科技）。Row 16 of the first batch (Moonshot AI).

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

  • Kimi K3Kimi K32026-06当前代current

    当前代。官方 README：2.8T 总参 MoE（Stable LatentMoE，896 专家激活 16），KDA + AttnRes 架构，原生多模态（文本/图像/视频），1M token 上下文；「world's first open 3T-class model」为厂商自评口径。Current generation. Official README: a 2.8T-parameter MoE (Stable LatentMoE, 16 of 896 experts active), KDA + AttnRes architecture, native multimodality (text/image/video), 1M-token context; "world's first open 3T-class model" is vendor-run wording.

    官方页面 ↗official page ↗
  • Kimi K2 系Kimi K2 line2025–2026

    上一代线：K2 / K2-Instruct / K2-Thinking / K2.5 / K2.6 / K2.7-Code 等在官方 org 在册（2025-07 起）。Previous line: K2 / K2-Instruct / K2-Thinking / K2.5 / K2.6 / K2.7-Code listed in the official org (from July 2025).

    官方页面 ↗official page ↗
  • Kimi-VL / Kimi-LinearKimi-VL / Kimi-Linear2025

    并行支线：视觉语言 Kimi-VL-A3B 与线性注意力 Kimi-Linear-48B-A3B（官方 org 在册）。Parallel branches: the vision-language Kimi-VL-A3B line and the linear-attention Kimi-Linear-48B-A3B (official org).

    官方页面 ↗official page ↗

尺寸档Size tiers

  • 旗舰 2.8T MoE（K3，激活数官方未逐条公开）Flagship 2.8T MoE (K3; active count not itemised officially)
  • 支线 7B–72B（Kimi-Audio / Kimi-Dev 等）Branch builds 7B–72B (Kimi-Audio / Kimi-Dev etc.)

许可Licences

  • otherKimi K3 License（模型卡 license_name: "kimi-k3"，非 OSI 标准许可，下载前读原文）Kimi K3 License (model-card license_name "kimi-k3"; not an OSI-standard licence — read the text before download) 许可原文 ↗licence text ↗

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

  • 官方 README：「a 2.8T-parameter model built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision capabilities and a 1-million-token context window」「activates 16 out of 896 experts — yielding an approximate 2.5× improvement in overall scaling efficiency over Kimi K2」（2.5× 为厂商自评口径）。Official README: "a 2.8T-parameter model built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision capabilities and a 1-million-token context window"; "activates 16 out of 896 experts — yielding an approximate 2.5× improvement in overall scaling efficiency over Kimi K2" (the 2.5× figure is vendor-run wording).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.