Skip to main content
AI Market Watch
Loading...

VIDRAFT

Category: Foundation Models / LLMs

Korean deep-tech AI startup building open-source foundation models (Darwin, AETHER) and an AI foundry for custom on-premise LLMs, plus autonomous scientific discovery platforms. VIDRAFT was founded in 2024. The company is led by 김민식 (Minsik Kim). Based in Seoul, South Korea. Team size: 11-50. Total funding raised: $480K. Latest round: Seed. Key investors include Undisclosed (seed round investors not publicly named).

Founded
2024
Headquarters
Seoul, South Korea
Team size
11-50
Total funding
$480K

Value proposition

Develops high-performance open-source reasoning models via proprietary recursive self-improvement (RSI) and model-merging technology, and delivers custom task-specialized LLMs (Darwin Factory) without GPU retraining or labeled data — enabling AI sovereignty and on-premise/closed-network deployment.

Products and solutions

Darwin model family (Darwin-180B-RSI, Darwin-397B-ZTC, Darwin-398B-JGOS), AETHER open-source foundation model (Aether-7B-5Attn, Apache 2.0), Darwin Factory AI foundry (diagnose-crossbreed-verify, custom LLM in one day), AX-RAY AI safety diagnostics (117 items), ZTC (Zero-Token Confidence), MARL inference middleware, POCKET on-device models (35B/26B), VIDOG autonomous quadruped robot kit, QuantumOS

Unique value

Recursive self-improvement (RSI) engine where models self-verify and retrain without human intervention, combined with selective model-merging (MRI-style diagnosis) that scales to 397B parameters at reduced inference cost; fully open weights and training recipes (AI sovereignty).

Target customer

Startups and mid-sized enterprises without in-house AI teams; defense, public sector, finance, and healthcare organizations requiring closed-network/on-premise AI where external cloud APIs are prohibited; research institutions.

Industries served

AI/Foundation Models, Drug Discovery & Biotech, Materials Science, Cybersecurity, Defense & Public Sector, Finance, Healthcare, Logistics, Robotics, Quantum Computing

Technology advantage

Proprietary RSI (recursive self-improvement) engine; selective model-merging/crossbreeding technology; ZTC zero-token confidence for pre-emptive hallucination control; MoE architecture (512 experts, 10 active); ~260K token context; Darwin-180B-RSI ranked #1 on 5 Hugging Face official leaderboards (AIME 2026 & HMMT 2026 perfect 100%, GPQA Diamond 94.44%, MMLU-Pro 88.12%, MMMU-Pro 79.48%); #1 on K-AI leaderboard; #1 in Google x Hugging Face Fast Gemma Challenge; 50+ official models, 400+ derivatives, 1M+ cumulative downloads.

How they differentiate

Claims 6x more official Hugging Face models and 7x+ cumulative downloads vs Sakana AI; differentiates RSI from data distillation by having models self-verify and retrain on validated solutions; offers fully open weights and training recipes (AI sovereignty) and closed-network/on-premise deployment for defense/public/finance where foreign APIs cannot enter.

Main competitors

Sakana AI (Japan, similar evolutionary/model-merging approach), Upstage (Korea, SOLAR LLM), Naver (HyperCLOVA X) / LG AI Research (EXAONE)

Key partnerships

Naver Cloud (cybersecurity-specialized AI foundation model consortium, MSIT national project, B200 x256 GPU support), Poba Biotech (AI drug candidate cell validation MOU), logistics fulfillment company (logistics-specialized 31B LLM & digital twin), Wearcube (strategic MOU & joint research), Ministry of SMEs and Startups 'Everyone's Challenge Bio' open innovation program, Seoul AI Hub resident, NIPA GPU support project.

Notable customers

NH Nonghyup Bank (PoC), Hwaseong City & NIA (civil complaint prediction RFP), K-defense cultivation program proposal, logistics fulfillment company (contract)

Major milestones

2024-03 founded, Darwin-180B-RSI #1 on 5 Hugging Face official leaderboards (Sep 2026), first perfect 100% on HF official math leaderboards (AIME/HMMT 2026), #1 K-AI leaderboard, #1 Google x Hugging Face Fast Gemma Challenge, selected for MSIT cybersecurity AI foundation model consortium with Naver Cloud (Sep 2026), selected for 'Everyone's Challenge Bio' open innovation (Sep 2026), Hugging Face Space of the Week (ai-world), 1M+ Darwin downloads.

Growth metrics

Darwin cumulative Hugging Face downloads 1M+ (4th Korean company/startup to reach this); POCKET on-device models 1.65M cumulative downloads (1.19M in last 30 days); Open Discovery Challenge 14,776+ drug candidates submitted; 50+ official models, 400+ derivatives.

Market positioning

Korean open-source foundation-model challenger positioning as an 'AI foundry' (analogous to TSMC for AI models) — selling the ability to grow customers' own AI rather than selling AI itself; targeting closed-network markets from 2027 and expanding to Japan and Middle East.

Geographic focus

South Korea (primary); expansion planned to Japan and Middle East

Patents and IP

9 patents held, including causal-leakage diagnosis method for neural network models (multi-axis diagnostic system) and N-(pyrrolidin-1-yl)-thiazol-4-yl-boronic acid compounds for progressive supranuclear palsy (PSP)/tauopathy treatment; 3 additional domestic patent applications for Darwin Family core technology.

About 김민식 (Minsik Kim)

Engineer/architect background with proven track record in global AI open-source ecosystem and community; technical CEO. Founded VIDRAFT in 2024.

Latest news about VIDRAFT

More Foundation Models / LLMs companies

Official website: