powered by ModelAtlas · 2,200+ models indexed

The DNA test
for AI models.

Identify model provenance, detect derivative lineage, and audit architecture before you deploy. Scan any HuggingFace model in seconds.

No backdoor detection. No security claims. Just architectural truth.

Scan a Model Free See How It Works

Stage 1 architecture screening is free and unlimited. No credit card required.

2,450+
Models in reference corpus
27
Base architecture families
<3s
Stage 1 screening latency
60+
Labs tracked
Labs covered in the ModelAtlas reference corpus
Meta · 661 models Qwen · 227 models Mistral · 212 models Google · 76 models Microsoft · 41 models DeepSeek · 29 models Zhipu AI · 23 models IBM · 19 models EleutherAI · 18 models OpenAI · 16 models Tencent · 12 models NVIDIA · 11 models MiniMax · 8 models Cohere · 8 models TII UAE · 9 models StepFun · 7 models AntGroup / inclusionAI · Bailing, Ling, LLaDA2 Moonshot · Kimi, Kimi-Linear, K2 ByteDance Seed · Seed-OSS Baidu · ERNIE 4.5, Qianfan Xiaomi · MiMo V2.x BAAI · Emu3, Emu3.5 OpenBMB · MiniCPM InternLM · S1 FreedomIntelligence · Pangu-R + 25 more labs
Why ModelDNA

More than a similarity score.

🧬

Deep Derivative Detection

We compare against 2,200+ verified architectures across 545 organizations. Identifies fine-tunes, distillations, and repackaged models that config checks miss.

Two-Stage Analysis

Stage 1 screens architecture configs instantly (free, unlimited). Stage 2 computes five weight-level signals for definitive lineage determination.

📋

Audit-Ready Output

Pro plans export SBOM-compatible reports and PDF audit trails with lineage origin, technique provenance, and compliance verdict — ready for governance review.

Live Example

See the analysis in action.

Stage 1 output is live today. Stage 2 weight analysis shown below is a roadmap preview.

$ modeldna scan poolside/Laguna-XS.2

[INIT]  Connecting to ModelAtlas reference database (2,241 models)
[SCAN]  Fetching config.json...

── STAGE 1: Architectural Screening ──────────────────────────
  Architecture:     LagunaForCausalLM
  Hidden size:      2048
  Vocab size:       100,352
  Lineage family:   GQA+SWA Hybrid
  Stage 1 verdict:  SIMILAR DESIGN → proceed to Stage 2

── STAGE 2: Weight-Level DNA Fingerprint ─────────────────────
  EAS  Embedding anchor similarity: 0.12  ▪▪░░░░░░░░  low
  END  Norm distribution:           0.31  ▪▪▪░░░░░░░  low
  NLF  Norm layer fingerprint:       0.28  ▪▪░░░░░░░░  low
  LEP  Layer energy profile:         0.41  ▪▪▪▪░░░░░░  moderate
  WVC  Weight cosine:                0.18  ▪░░░░░░░░░  low

── VERDICT ────────────────────────────────────────────────────
  Composite DNA Score:  0.26 / 1.00
  Classification:       ARCHITECTURAL INSPIRATION (not weight-derived)
  Lineage:              GQA+SWA Hybrid — adopted pattern from StepFun (Feb 2026)
  Weights:              INDEPENDENT — different vocab rules out tokenizer inheritance

  ✓ No weight inheritance detected
Report saved: ./modeldna-report-laguna-xs2.json
Powered by ModelAtlas · modeldna.ai · a RadicalNotion product
Process

Two stages. One verdict.

1

Architecture Screening

We parse the model config to identify architecture family, vocab, and technique signatures. Resolves the majority of cases instantly.

FREE · UNLIMITED · <3s
2

Weight-Level Analysis ROADMAP

Five statistical signals (EAS, END, NLF, LEP, WVC) will analyze weight distributions to detect derivative relationships configs can't reveal.

STAGE 2 · Coming soon
3

Enriched Verdict

Composite DNA score, lineage classification, and optional ModelAtlas enrichment — VRAM estimates, ARS scores, and technique origin chain.

PRO · JSON / PDF / SBOM
Pricing

Developer-priced. Enterprise-capable.

The ModelAtlas knowledge engine runs on RadicalNotion infrastructure. You get enterprise-grade depth at a fraction of incumbent pricing.

All scanning is Stage 1 — architecture screening from config.json only. No weight download required. Stage 2 weight-level analysis is on the roadmap.

Free
$0/mo

For anyone using open weights on HuggingFace.

  • Unlimited Stage 1 scanning — all public HF models
  • Architecture family identification
  • Claim validation — flags unverifiable assertions
  • Derivative discovery — find models sharing your base
  • JSON output
Start Scanning
Custom
Contact Us

For open weight hosters, compliance teams, and enterprises.

  • Everything in Pro
  • Private model ingestion + on-premise deployment
  • Custom ModelAtlas reference databases
  • Automated moderation for model hosting platforms
  • SLA guarantees + dedicated engineering
Contact Sales
Honest Answers

What modeldna does — and doesn't.

Does this detect backdoors or malware?

No. modeldna analyzes statistical fingerprints for provenance and derivative lineage. We do not perform security vulnerability scanning or malware detection. These are different problems requiring different tools.

Is this cryptographic proof of ownership?

No. Our results are based on statistical fingerprinting — high-confidence evidence of lineage, not cryptographic signatures. Results are suitable for compliance review and audit documentation but not legal disputes over IP ownership.

How does the ModelAtlas reference database work?

ModelAtlas continuously scans and indexes public model releases, extracting architecture signatures, technique taxonomies, and lineage relationships. The reference database is private and updated weekly. The CLI queries it via API — you never see the raw database.

Can I scan quantized models (GGUF, AWQ, FP8)?

Stage 1 works on any model with a config.json. Stage 2 weight analysis requires dequantization to BF16 first — this is handled automatically for supported formats. GGUF Q4/Q8 and safetensors FP8 are supported; AWQ/GPTQ support is in development.