VIDRAFT Surpasses 1.6 Million Hugging Face Downloads: Darwin, AETHER, and the AI Foundry Approach Explained
TL;DR: Korean Pre-AGI startup VIDRAFT has crossed 1.6 million cumulative model downloads on Hugging Face, placing it among just four Korean organizations to exceed 1 million downloads. Their self-evolving Darwin model family leads multiple public benchmarks — including GPQA Diamond and the Polaris drug-prediction blind benchmark — while a from-scratch architecture called AETHER signals their push toward proprietary model design. Developers can explore their models and safety tooling directly on Hugging Face.
What it is
VIDRAFT (비드래프트, CEO Kim Min-sik) is a Korean Pre-AGI AI startup that develops and publishes large language and science-reasoning models. As of September 1, 2026, their cumulative Hugging Face model downloads have exceeded 1.6 million — a milestone that puts them alongside LG, Kakao, and Upstage as the only four Korean organizations to break the 1 million download threshold on the platform.
Their public model portfolio spans several distinct lines:
- Darwin series — A self-evolving model family optimized for scientific reasoning and pharmaceutical prediction, including the published Darwin-398B-JGOS variant.
- AETHER — A from-scratch foundation model (including the published Aether-7B-5Attn) representing VIDRAFT's independent architecture research track, separate from fine-tuning or adaptation of existing base models.
- AX-RAY — An AI safety vulnerability diagnostic tool and accompanying evaluation datasets, publicly released on Hugging Face.
- BUILDING (빌딩) — A free-to-access generative AI tool for architectural design.
The company also participates in Korean government AI initiatives and enterprise on-premises AI deployments.
How it works
VIDRAFT describes its core competency as an "AI Foundry process" — a manufacturing-style capability for building, adapting, and hardening AI models for specific domains and customers, rather than producing a single flagship model.
At a conceptual level, their approach involves two complementary tracks:
Breeding and knowledge grafting (Darwin · Chimera): Existing models are cross-trained and domain-adapted using specialized data — including MRI-based machine learning for medical diagnostics and prediction — to improve capability in targeted verticals. The "Chimera" technique name suggests structured hybridization of model capabilities across domains.
From-scratch architecture research (AETHER): In parallel, VIDRAFT develops models built entirely on their own architecture, allowing them to accumulate independent design expertise rather than remaining dependent on external base model releases.
Safety is a first-class engineering concern: AX-RAY is their internal-turned-public diagnostic pipeline for identifying AI safety vulnerabilities, which can inform deployment hardening for enterprise customers.
Their stated product roadmap extends these foundations toward:
- Enterprise on-premises small LLMs (sLLM) with reduced inference cost
- Pocket on-device AI models
- Vertically specialized operating-system-style platforms: NationalOS, PharmaOS, and MaterialsOS (the last integrating generative AI with IBM quantum computer validation for novel materials discovery)
Benchmarks & results
All figures below come directly from the ZDNet Korea source article and refer to public leaderboards and externally verified competitions:
| Benchmark / Competition | Result |
|---|---|
| GPQA Diamond (graduate-level science reasoning) | 90.9% — Darwin-398B-JGOS, ranked #1 on public leaderboard |
| Polaris (drug prediction external blind benchmark) | #1 in 16 categories |
| K-AI Leaderboard (operated by Korea MSIT & NIA) | #1 ranking maintained |
| Fast Gemma Challenge (Google × Hugging Face global competition) | Top record achieved |
CEO Kim Min-sik's stated philosophy on these results: "Performance must be proven by public metrics and verification by the organizers, not by claims."
How to try it
VIDRAFT's models and tooling are publicly available on Hugging Face. You can browse their organization page and download models using standard tooling:
# Install the Hugging Face CLI
pip install huggingface_hub
# Browse and download VIDRAFT models (replace <model-name> with the actual repo name)
huggingface-cli download vidraft/<model-name>
Note: Specific model repository names and access conditions vary per release. Check the VIDRAFT Hugging Face organization page for current availability, licensing, and any gated-access requirements. AX-RAY evaluation datasets are also published there.
No public OpenAI-compatible API endpoint or GitHub organization URL was cited in the source article at time of writing.
FAQ
Q: Is Darwin-398B-JGOS a fine-tune of an existing open model, or a novel architecture?
A: The source article does not explicitly specify the base architecture. The Darwin line is described as "self-evolving" and uses cross-breeding and knowledge grafting techniques; AETHER (e.g., Aether-7B-5Attn) is explicitly the from-scratch architecture track. Check the model cards on Hugging Face for lineage details.
Q: What is AX-RAY, and can I use it to evaluate my own models?
A: AX-RAY is VIDRAFT's AI safety vulnerability diagnostic tool, released publicly on Hugging Face along with its evaluation datasets. It is designed to surface safety weaknesses in language models. The public release suggests it can be applied to third-party models, but refer to the Hugging Face repository's documentation and license for specifics on scope and usage terms.
Originally reported by ZDNet Korea (2026-09-01) — source article.
Top comments (0)