{
  "story_id": "10e9a186f969cbff7d46626cf2aefcd9",
  "desk": "drm3",
  "revision": 2,
  "published_at": "2026-09-03T13:13:48.000Z",
  "content_hash": "9325a9b12fa97b66f865ca02b154c523da395a0947c4982965cdb00453f98e5e",
  "hash_basis": "sha256 over `headline\\ndek\\nprose`, plus `\\n` + the canonical citations JSON when any source is placed, plus `\\n#blog` for blogs",
  "basis": {
    "headline": "NeoMME: Efficient Multilingual Multimodal Encoder for Visual Document Retrieval",
    "dek": "NeoMME introduces a single Transformer encoder for multilingual text and images, achieving state-of-the-art visual document retrieval.",
    "prose": "NeoMME processes both text tokens and raw image patches using a single bidirectional Transformer without a separate pretrained vision tower or causal language model. [^1]\n\nThe authors introduce NeoMME, a family of multilingual multimodal encoders consisting of 260M and 800M parameter variants. [^2] NeoMME-Retriever was fine-tuned for visual document retrieval using the page-image approach introduced by ColPali. [^3] Both NeoMME models are pretrained from scratch with a masked discrete-diffusion text objective, conditioned on visible image patches for multimodal examples. [^4][^5]\n\nOn the ViDoRe v3 benchmark, the resulting NeoMME-Retriever 260M outperforms all evaluated models strictly below 800M parameters with an nDCG@10 score of 0.523. [^6]",
    "cited": "[{\"statement\":\"NeoMME processes both text tokens and raw image patches using a single bidirectional Transformer without a separate pretrained vision tower or causal language model.\",\"source\":\"Hugging Face\",\"instrument\":\"News\",\"claim_key\":null,\"published_at\":\"2026-09-03T13:13:48.000Z\",\"publisher_count\":1,\"sources\":[\"Hugging Face\"]},{\"statement\":\"The authors introduce NeoMME, a family of multilingual multimodal encoders consisting of 260M and 800M parameter variants.\",\"source\":\"Hugging Face\",\"instrument\":\"News\",\"claim_key\":null,\"published_at\":\"2026-09-03T13:13:48.000Z\",\"publisher_count\":1,\"sources\":[\"Hugging Face\"]},{\"statement\":\"NeoMME-Retriever was fine-tuned for visual document retrieval using the page-image approach introduced by ColPali.\",\"source\":\"Hugging Face\",\"instrument\":\"News\",\"claim_key\":null,\"published_at\":\"2026-09-03T13:13:48.000Z\",\"publisher_count\":1,\"sources\":[\"Hugging Face\"]},{\"statement\":\"The entire NeoMME model is trained from scratch using a masked discrete-diffusion objective.\",\"source\":\"Hugging Face\",\"instrument\":\"News\",\"claim_key\":null,\"published_at\":\"2026-09-03T13:13:48.000Z\",\"publisher_count\":1,\"sources\":[\"Hugging Face\"]},{\"statement\":\"Both NeoMME models are pretrained from scratch with a masked discrete-diffusion text objective, conditioned on visible image patches for multimodal examples.\",\"source\":\"arXiv.org\",\"instrument\":\"News\",\"claim_key\":null,\"published_at\":\"2026-09-03T04:00:00.000Z\",\"publisher_count\":1,\"sources\":[\"arXiv.org\"]},{\"statement\":\"On the ViDoRe v3 benchmark, the resulting NeoMME-Retriever 260M outperforms all evaluated models strictly below 800M parameters with an nDCG@10 score of 0.523.\",\"source\":\"arXiv.org\",\"instrument\":\"News\",\"claim_key\":null,\"published_at\":\"2026-09-03T04:00:00.000Z\",\"publisher_count\":1,\"sources\":[\"arXiv.org\"]}]",
    "kind": "news"
  },
  "receipt_verify": "Ed25519 over the dot-joined string `slice_hash.cursor_from.cursor_to.view.view_version.row_count`; public_key and sig are base64url of the raw 32-byte key / 64-byte signature",
  "receipt": null,
  "receipt_note": "this revision predates receipt-keeping (before v0.37.0); the filed row lives in the record",
  "generation_chain": {
    "wire": {
      "stream": "fountain_news",
      "story_id": "98a1c2ba0084257df42487946f40edf2",
      "thread_id": "5c6835aa75ae4b9449818592a0259959",
      "thread_label": "ColModernVBERT",
      "novelty": "UPDATE",
      "content_hash": "152b953721a5874e88f13017e698d80d7c61808b01c8ef6467260903f36d8a25",
      "last_published_at": "2026-09-03T13:13:48.000Z",
      "read_receipt": {
        "slice_hash": "9d748ae04a83802cd61b786e1385a790409ae163dd0c42fe240fe49962bef9e3",
        "cursor_from": "eyJ0cyI6IjIwMjYtMDktMDNUMTM6MDA6MDAuMDAwMDAwWiIsImlkIjoiNDQxNWY3MmVjNjI3YWFmZmYyOWNkMjU2NGE5MWIxMDAiLCJ2IjoiMSJ9",
        "cursor_to": "eyJ0cyI6IjIwMjYtMDktMDNUMTM6NDQ6MjMuNzk4MDAwWiIsImlkIjoiNGUyYWVlOGQzZmY4OGNhZTBlMjMwOWQ4OGIzZjQyMDEiLCJ2IjoiMSJ9",
        "view": "v_fountain_news",
        "view_version": "1",
        "row_count": 100,
        "window_days": 3,
        "bytes_scanned": 12166856,
        "credits": 8,
        "price_per_100_rows": 8,
        "sig": "IXw5gtyyWIY-4A9RlHJYzWzRJSy_qiJLoaWrchZRrVUztlMLqy94akiAi_UnMWDVS8_jzcYPkpn4zZqx6ILpBA",
        "public_key": "bMUigy8O0jOnBxQ4Sc-5lwhIZ8LQVAhxMbR7qESVuUE",
        "signer_path": "lakehouse/data-extract/v1",
        "alg": "Ed25519",
        "signed": true
      },
      "compose": 3,
      "recomposed_at": "2026-09-05T20:55:43.784Z"
    },
    "written_at": "2026-09-04T06:19:35.265Z"
  },
  "cited_facts": [
    {
      "n": 1,
      "statement": "NeoMME processes both text tokens and raw image patches using a single bidirectional Transformer without a separate pretrained vision tower or causal language model.",
      "source": "Hugging Face",
      "instrument": "News",
      "claim_key": null,
      "url": null,
      "occurred_at": null,
      "published_at": null
    },
    {
      "n": 2,
      "statement": "The authors introduce NeoMME, a family of multilingual multimodal encoders consisting of 260M and 800M parameter variants.",
      "source": "Hugging Face",
      "instrument": "News",
      "claim_key": null,
      "url": null,
      "occurred_at": null,
      "published_at": null
    },
    {
      "n": 3,
      "statement": "NeoMME-Retriever was fine-tuned for visual document retrieval using the page-image approach introduced by ColPali.",
      "source": "Hugging Face",
      "instrument": "News",
      "claim_key": null,
      "url": null,
      "occurred_at": null,
      "published_at": null
    },
    {
      "n": 4,
      "statement": "The entire NeoMME model is trained from scratch using a masked discrete-diffusion objective.",
      "source": "Hugging Face",
      "instrument": "News",
      "claim_key": null,
      "url": null,
      "occurred_at": null,
      "published_at": null
    },
    {
      "n": 5,
      "statement": "Both NeoMME models are pretrained from scratch with a masked discrete-diffusion text objective, conditioned on visible image patches for multimodal examples.",
      "source": "arXiv.org",
      "instrument": "News",
      "claim_key": null,
      "url": null,
      "occurred_at": null,
      "published_at": null
    },
    {
      "n": 6,
      "statement": "On the ViDoRe v3 benchmark, the resulting NeoMME-Retriever 260M outperforms all evaluated models strictly below 800M parameters with an nDCG@10 score of 0.523.",
      "source": "arXiv.org",
      "instrument": "News",
      "claim_key": null,
      "url": null,
      "occurred_at": null,
      "published_at": null
    }
  ],
  "note": "A signature proves who filed this and that it has not changed since. It never makes a claim true. The dates on cited_facts (occurred_at, published_at) are the forge's, joined at read time; they are not part of the signed basis above."
}