---
title: Model Genome Korea
emoji: 🧬
colorFrom: green
colorTo: blue
sdk: static
pinned: true
license: apache-2.0
short_description: DNA lineage test of Korean LLM & VLM foundation models
tags:
  - korean-llm
  - korean-ai
  - model-lineage
  - model-provenance
  - foundation-models
  - vlm
  - sovereign-ai
models:
  - LGAI-EXAONE/EXAONE-3.5-7.8B-Instruct
  - LGAI-EXAONE/EXAONE-4.0-32B
  - LGAI-EXAONE/EXAONE-4.5-33B
  - LGAI-EXAONE/K-EXAONE-236B-A23B
  - LGAI-EXAONE/K-EXAONE-2.0-750B-A37B
  - upstage/SOLAR-10.7B-v1.0
  - upstage/Solar-Open-100B
  - upstage/Solar-Open2-250B
  - skt/A.X-4.0-Light
  - skt/A.X-4.0
  - skt/A.X-K2
  - skt/A.X-4.0-VL-Light
  - skt/A.X-3.1
  - NCSOFT/VARCO-VISION-2.0-14B
  - NCSOFT/VARCO-VISION-2.0-1.7B
  - NCSOFT/Llama-VARCO-8B-Instruct
  - naver-hyperclovax/HyperCLOVAX-SEED-Think-32B
  - naver-hyperclovax/HyperCLOVAX-SEED-Think-14B
  - naver-hyperclovax/HyperCLOVAX-SEED-Vision-Instruct-3B
  - naver-hyperclovax/HyperCLOVAX-SEED-Omni-8B
  - kakaocorp/kanana-1.5-8b-instruct-2505
  - kakaocorp/kanana-2-30b-a3b-instruct-2601
  - kakaocorp/kanana-2-3b-instruct
  - kakaocorp/kanana-1.5-v-3b-instruct
  - kakaocorp/kanana-1.5-15.7b-a3b-instruct
  - K-intelligence/Midm-2.0-Base-Instruct
  - K-intelligence/Midm-2.0-Mini-Instruct
  - Motif-Technologies/Motif-3-Beta
  - Motif-Technologies/Motif-2-12.7B-Base
  - Motif-Technologies/Motif-2.6B
  - FINAL-Bench/Aether-7B-5Attn
  - FINAL-Bench/AETHER-7B-7Attn-base
  - FINAL-Bench/Aether-6B-11Attn-base
---

# Model Genome Korea — Architecture DNA Test of Korean LLM & VLM Foundation Models

**Model Genome Korea is a free, interactive tool that reveals the architectural lineage ("DNA") of every major Korean sovereign AI foundation model.** It genotypes each Korean large language model (LLM) and vision-language model (VLM) — measured directly from public `config.json` files and model weights on Hugging Face — to show, objectively, **which Korean AI models are built from scratch and which are derived from foreign open-weight bases like Qwen, Llama, or DeepSeek.**

The tool covers foundation models from **9 Korean organizations** that released self-developed foundation models on Hugging Face: **LG AI Research (EXAONE), SKT (A.X), Upstage (SOLAR), NCSOFT (VARCO), NAVER (HyperCLOVA X), Kakao (Kanana), KT (Mi:dm), Motif Technologies (Motif), and VIDRAFT (AETHER)** — spanning large enterprises, telcos, mid-size firms, and startups.

## What is Model Genome Korea?

Model Genome Korea is an architecture-provenance explorer for Korean sovereign AI. Like a genetic paternity test for neural networks, it fingerprints each model's architecture and weights and reports a **genotype** — a single, at-a-glance label describing how original (or derived) the model is. Every value is measured from public artifacts, so the results are reproducible and vendor-neutral. This is a transparency tool, **not an accusation**: building on open-weight foundations (Qwen, Llama, DeepSeek, Mistral) is a standard and legitimate engineering practice across the industry.

## How it works: the two-axis genotype

Each model is scored on two independent axes, both measured objectively:

1. **Architecture** — Does the `model_type` and the shape signature (hidden size · intermediate size · layer count) match a known foreign open-weight model? An exact match is strong evidence of an adopted architecture.
2. **Weights** — Embedding **CKA (Centered Kernel Alignment)**, which is rotation-invariant, distinguishes *from-scratch* pretraining from *continued-pretraining* on a foreign base.

A third dimension, **attention diversity** (the number of distinct attention mechanisms), is reported as a measure of architectural originality.

## Genotypes explained

| Genotype | Meaning | Example models |
|---|---|---|
| 🟢 **Native** | Self-developed architecture **and** from-scratch weights | EXAONE-4.0, K-EXAONE-2.0, Solar-Open2, Motif, AETHER |
| 🔵 **Adapted** | Mostly self, one axis borrowed | EXAONE-3.5, A.X-3.1 |
| 🟡 **Mixed** | Partial inheritance from foreign bases | Mi:dm 2.0, HyperCLOVA X, Kanana-2 |
| 🔴 **Ported** | Foreign architecture **and** foreign weights | VARCO-VISION (Qwen3), Kanana-1.5 (Llama), A.X-4.0 (Qwen2.5) |

## Korean AI companies and models covered

| Company | Series | Representative model | HF org |
|---|---|---|---|
| LG AI Research | EXAONE / K-EXAONE | K-EXAONE-2.0-750B | `LGAI-EXAONE` |
| SKT | A.X (A dot X) | A.X-K2 | `skt` |
| Upstage | SOLAR / Solar-Open | Solar-Open2-250B | `upstage` |
| NCSOFT | VARCO | VARCO-VISION-2.0 | `NCSOFT` |
| NAVER | HyperCLOVA X SEED | HyperCLOVAX-SEED-Think-32B | `naver-hyperclovax` |
| Kakao | Kanana | kanana-2-30b-a3b | `kakaocorp` |
| KT | Mi:dm 2.0 | Midm-2.0-Base | `K-intelligence` |
| Motif Technologies | Motif | Motif-3-Beta | `Motif-Technologies` |
| VIDRAFT | AETHER | AETHER-7B-5Attn | `FINAL-Bench` |

## Key findings

- **SKT A.X-4.0** exactly matches the **Qwen2.5** architecture (hidden 3584 · intermediate 18944 · 28 layers); A.X-K2 matches **DeepSeek-V3**.
- **NCSOFT VARCO-VISION 2.0** is built on **Qwen3** (14B and 1.7B), with the base model disclosed on its card.
- **Kakao Kanana-1.5-8B** matches **Llama-3.1-8B** exactly; Kanana-2 moved to a self tokenizer.
- **LG EXAONE** uses a self architecture family and self tokenizer; **K-EXAONE-2.0-750B** is released under **Apache-2.0**.
- **Motif** and **Upstage Solar-Open2** use fully self-developed MoE architectures with no foreign match.
- **AETHER** uses heterogeneous attention (mamba2, hyena, MLA, linear, GDN, NSA, sliding) with from-scratch weights.

## Frequently asked questions

### Are Korean LLMs built from scratch or based on Qwen and Llama?
Both. Many Korean commercial LLMs adopt a foreign open-weight foundation: SKT A.X-4.0 uses the Qwen2.5 architecture, NCSOFT VARCO-VISION is built on Qwen3, and Kakao Kanana-1.5 is built on Llama-3.1. Models with fully self-developed architecture **and** from-scratch weights include Motif, Upstage Solar-Open2, LG EXAONE-4.0, and AETHER.

### Which Korean AI model has the most original architecture?
Measured by the number of distinct attention mechanisms (a proxy for architectural originality), AETHER leads with heterogeneous attention combining mamba2, hyena, MLA, linear attention, gated delta net, native sparse attention, and sliding-window attention. LG EXAONE-4.0 and Motif-3-Beta follow with hybrid attention. Most other Korean models use a single standard attention (grouped-query attention or multi-head latent attention).

### Is LG EXAONE open source?
The EXAONE 3.5/4.0/4.5 series is released under a custom EXAONE AI license (`license: other`). However, the latest flagship, **K-EXAONE-2.0-750B, is released under Apache-2.0** — a fully open-source license.

### Which Korean AI models use a non-commercial license?
NCSOFT's VARCO-VISION models are released under **CC-BY-NC-4.0** (non-commercial). By contrast, SKT A.X, KT Mi:dm 2.0, Kakao Kanana-1.5-8B, Motif-2, LG K-EXAONE-2.0, and AETHER are released under permissive Apache-2.0 or MIT licenses.

### Is SKT A.X a Qwen model?
SKT A.X-4.0-Light and A.X-4.0 share the exact shape signature of Qwen2.5-7B and Qwen2.5-72B respectively, with a Korean-adapted tokenizer. A.X-K2 shares the DeepSeek-V3 shape signature. The architecture is adopted; the models are then further trained for Korean.

### What does "model genotype" mean?
It is a single label — Native, Adapted, Mixed, or Ported — summarizing how much of a model's architecture and weights are self-developed versus inherited from a foreign open-weight base, computed from public configuration files and embedding similarity.

## Methodology and reproducibility

- **Architecture axis:** `model_type` plus (hidden_size · intermediate_size · num_hidden_layers) compared against foreign open-weight references (Qwen2.5, Qwen3, Llama-3, DeepSeek-V3, Mistral).
- **Weights axis:** embedding-matrix CKA (rotation-invariant) between a model and its candidate base; near-zero CKA indicates independent, from-scratch pretraining.
- **Attention diversity:** count of distinct attention mechanisms declared in `config.json` (e.g. layer_types, linear_attn_config, mamba2/hyena/MLA blocks).

All inputs are public. The same yardstick is applied to every model without exception.

## Data sources

Public Hugging Face model repositories: `config.json`, `tokenizer.json`, and safetensors embedding tensors, measured 2026-08. The interface defaults to English and switches to Korean or Chinese based on the visitor's browser language.

---

*Model Genome Korea is an independent, vendor-neutral transparency project. Model names, companies, and licenses are the property of their respective owners.*