{
  "slug": "kimi-linear-48b-a3b",
  "order": 24,
  "name": "Kimi Linear (48B-A3B)",
  "baseTitle": "Kimi Linear 48B-A3B",
  "company": "Moonshot AI",
  "country": "China",
  "flag": "🇨🇳",
  "date": "2025-10-30",
  "decoderType": "Sparse hybrid",
  "attention": "3:1 Kimi Delta Attention and MLA",
  "layerMix": "7 MLA + 20 Kimi Delta Attention",
  "scale": "48B total, 3B active (6.3% active)",
  "totalParams": 48000000000,
  "activeParams": 3000000000,
  "context": "1,000,000",
  "contextTokens": 1000000,
  "license": "MIT License",
  "kvCacheText": "7.9 KiB",
  "kvCacheKiB": 7.9,
  "kvCacheLevel": "Very low",
  "summary": "Linear-attention hybrid that keeps a transformer backbone but replaces most full-attention layers.",
  "keyDetail": "Uses NoPE in MLA layers, channel-wise KDA gating, and kernel-4 ShortConv preprocessing for Q, K, and V in KDA layers.",
  "image": {
    "full": "https://sebastianraschka.com/llm-architecture-gallery/images/architectures/kimi-linear-48b-a3b.webp",
    "thumb": "https://sebastianraschka.com/llm-architecture-gallery/images/architectures/thumbnails/kimi-linear-48b-a3b.webp",
    "alt": "Figure 42: Kimi Linear 48B-A3B",
    "width": 960,
    "height": 883
  },
  "articleUrl": "https://magazine.sebastianraschka.com/p/the-big-llm-architecture-comparison#%C2%A7144-kimi-linear-vs-qwen3-next",
  "links": [
    {
      "label": "View in article",
      "href": "https://magazine.sebastianraschka.com/p/the-big-llm-architecture-comparison#%C2%A7144-kimi-linear-vs-qwen3-next",
      "title": null
    },
    {
      "label": "config.json",
      "href": "https://huggingface.co/moonshotai/Kimi-Linear-48B-A3B-Base/blob/main/config.json",
      "title": "moonshotai/Kimi-Linear-48B-A3B-Base"
    },
    {
      "label": "Tech report",
      "href": "https://arxiv.org/pdf/2510.26692",
      "title": null
    }
  ],
  "concepts": [
    {
      "slug": "nope",
      "label": "NoPE",
      "name": "No Positional Embeddings",
      "url": "https://llm-atlas-7or.pages.dev/concepts/nope/"
    },
    {
      "slug": "mla",
      "label": "MLA",
      "name": "Multi-Head Latent Attention",
      "url": "https://llm-atlas-7or.pages.dev/concepts/mla/"
    },
    {
      "slug": "hybrid-attention",
      "label": "Hybrid Attention",
      "name": "Hybrid Attention",
      "url": "https://llm-atlas-7or.pages.dev/concepts/hybrid-attention/"
    },
    {
      "slug": "shortconv",
      "label": "ShortConv",
      "name": "Short Convolution",
      "url": "https://llm-atlas-7or.pages.dev/concepts/shortconv/"
    },
    {
      "slug": "attention-residuals",
      "label": "Attention Residuals",
      "name": "Attention Residuals",
      "url": "https://llm-atlas-7or.pages.dev/concepts/attention-residuals/"
    }
  ],
  "aai": {
    "v4": {
      "label": "AA Intelligence Index v4.3.2",
      "total": 7.3,
      "profile": {},
      "url": "https://artificialanalysis.ai/models/kimi-linear-48b-a3b-instruct"
    }
  },
  "sourceUrl": "https://sebastianraschka.com/llm-architecture-gallery/#card-kimi-linear-48b-a3b",
  "family": "hybrid-moe",
  "url": "https://llm-atlas-7or.pages.dev/models/kimi-linear-48b-a3b/",
  "markdownUrl": "https://llm-atlas-7or.pages.dev/models/kimi-linear-48b-a3b.md"
}