Documentation

Scan a model, read the report.

ModelDNA Stage 1 runs on a public HuggingFace Space. There's no install, no signup, and no SDK to learn. Paste a model ID, wait a few seconds, copy the JSON.

Try it

The scanner runs on a HuggingFace Space. HuggingFace blocks iframe embedding for security, so it opens in a new tab.

huggingface.co/spaces
RadicalNotionAI / modeldna

Paste any public HuggingFace model ID, get a Stage 1 verdict in a few seconds. Free, no signup.

What you send

A HuggingFace model identifier. That's it. Formats accepted:

  • <org>/<model> — e.g. meta-llama/Llama-3.2-1B
  • Full URL — e.g. https://huggingface.co/Qwen/Qwen3-4B

The scanner fetches the model's config.json (a few KB), compares it against the ModelAtlas reference corpus, and emits a verdict. Weights are not downloaded for Stage 1.

What you get back

A JSON report. Below is an annotated example; field names are stable and safe to depend on.

{
  "model": "poolside/Laguna-XS.2",
  "scanned_at": "2026-05-25T17:08:40Z",
  "stage": 1,

  // architectural fingerprint pulled straight from config.json
  "architecture": {
    "class":        "LagunaForCausalLM",
    "hidden_size":  2048,
    "vocab_size":   100352,
    "family":       "GQA+SWA Hybrid"
  },

  // closest matches in the ModelAtlas reference corpus
  "matches": [
    { "model": "StepFun/Step-Audio-2", "score": 0.82 },
    { "model": "Qwen/Qwen3-4B",         "score": 0.71 }
  ],

  // verdict — the field you should read first
  "verdict": {
    "classification": "ARCHITECTURAL_INSPIRATION",
    "confidence":     "high",
    "lineage":        "GQA+SWA Hybrid — pattern adopted from StepFun (Feb 2026)",
    "next_step":      "stage_2_recommended"
  }
}

Classification values

  • INDEPENDENT — no meaningful architectural similarity to known models.
  • ARCHITECTURAL_INSPIRATION — adopts a recognizable design pattern (GQA, SWA, MoE routing shape, etc.) without weight inheritance.
  • DERIVATIVE_CANDIDATE — config is close enough to a known model that Stage 2 weight analysis is recommended before drawing conclusions.
  • CONFIRMED_DERIVATIVE — Stage 2 only. Weight-level signals indicate the model is fine-tuned, distilled, or repackaged from a specific upstream.

What it can't tell you

  • Whether the model contains backdoors or malicious weights. That's a different problem with different tooling.
  • Cryptographic proof of ownership. Lineage scores are statistical, not legal evidence.
  • Anything about private models you haven't made public on HuggingFace. Contact us for on-premise deployment (tim@radicalnotion.ai).

REST API

ROADMAP

A versioned POST /v1/scan endpoint ships with the Pro plan. Same JSON shape as above, plus bulk scanning and webhook notifications when known derivatives of your models appear on HuggingFace. Email tim@radicalnotion.ai for early access.