{
  "slug": "deepseek-v4-flash",
  "order": 60,
  "name": "DeepSeek V4-Flash (284B)",
  "baseTitle": "DeepSeek V4-Flash",
  "company": "DeepSeek",
  "country": "China",
  "flag": "🇨🇳",
  "date": "2026-04-24",
  "decoderType": "Sparse MoE",
  "attention": "MLA-style CSA/HCA with mHC",
  "layerMix": "43 CSA/HCA",
  "scale": "284B total, 13B active (4.6% active)",
  "totalParams": 284000000000,
  "activeParams": 13000000000,
  "context": "1,048,576",
  "contextTokens": 1048576,
  "license": "MIT License",
  "kvCacheText": "5.4 KiB",
  "kvCacheKiB": 5.4,
  "kvCacheLevel": "Very low",
  "summary": "DeepSeek's efficient V4 preview keeps the million-token architecture while reducing the MoE scale to 284B parameters and 13B active parameters.",
  "keyDetail": "Uses 256 experts, 6 routed plus 1 shared expert per token, hash-based routing, compressed attention, and the V4 MTP path.",
  "image": {
    "full": "https://sebastianraschka.com/llm-architecture-gallery/images/architectures/deepseek-v4-flash.webp",
    "thumb": "https://sebastianraschka.com/llm-architecture-gallery/images/architectures/thumbnails/deepseek-v4-flash.webp",
    "alt": "Figure 17: DeepSeek V4-Flash",
    "width": 960,
    "height": 774
  },
  "articleUrl": "https://magazine.sebastianraschka.com/p/recent-developments-in-llm-architectures#%C2%A75-csahca-mhc-and-compressed-attention-caches-deepseek-v4",
  "links": [
    {
      "label": "View in article",
      "href": "https://magazine.sebastianraschka.com/p/recent-developments-in-llm-architectures#%C2%A75-csahca-mhc-and-compressed-attention-caches-deepseek-v4",
      "title": null
    },
    {
      "label": "config.json",
      "href": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/blob/main/config.json",
      "title": "deepseek-ai/DeepSeek-V4-Flash"
    },
    {
      "label": "License",
      "href": "https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash/blob/main/LICENSE",
      "title": null
    },
    {
      "label": "Tech report",
      "href": "https://arxiv.org/abs/2606.19348",
      "title": null
    }
  ],
  "concepts": [
    {
      "slug": "csa-hca",
      "label": "CSA/HCA",
      "name": "Compressed Sparse & Heavily Compressed Attention",
      "url": "https://llm-atlas-7or.pages.dev/concepts/csa-hca/"
    },
    {
      "slug": "mhc",
      "label": "mHC",
      "name": "Manifold-constrained Hyper-Connections",
      "url": "https://llm-atlas-7or.pages.dev/concepts/mhc/"
    },
    {
      "slug": "mla",
      "label": "MLA",
      "name": "Multi-Head Latent Attention",
      "url": "https://llm-atlas-7or.pages.dev/concepts/mla/"
    },
    {
      "slug": "deepseek-sparse-attention",
      "label": "DSA",
      "name": "DeepSeek Sparse Attention",
      "url": "https://llm-atlas-7or.pages.dev/concepts/deepseek-sparse-attention/"
    },
    {
      "slug": "moe",
      "label": "MoE",
      "name": "Mixture of Experts",
      "url": "https://llm-atlas-7or.pages.dev/concepts/moe/"
    },
    {
      "slug": "mtp",
      "label": "MTP",
      "name": "Multi-Token Prediction",
      "url": "https://llm-atlas-7or.pages.dev/concepts/mtp/"
    }
  ],
  "aai": {
    "v4": {
      "label": "AA Intelligence Index v4.3.2",
      "total": 34.3,
      "profile": {
        "General": 31.8,
        "Scientific": 27.6,
        "Coding": 31.2,
        "Agents": 43.5
      },
      "url": "https://artificialanalysis.ai/models/deepseek-v4-flash"
    }
  },
  "sourceUrl": "https://sebastianraschka.com/llm-architecture-gallery/#card-deepseek-v4-flash",
  "family": "moe",
  "url": "https://llm-atlas-7or.pages.dev/models/deepseek-v4-flash/",
  "markdownUrl": "https://llm-atlas-7or.pages.dev/models/deepseek-v4-flash.md"
}