A Seoul-based AI startup raises the bar for transparency and reproducibility in the Korean AI ecosystem.
TL;DR: Korean AI startup VIDRAFT has released Aether-7B-5Attn, a fully open-source foundation model, on Hugging Face under the Apache-2.0 license. Unlike most "open-source" LLMs that share only model weights, VIDRAFT has made available the training data recipe, training code, hyperparameters, training logs, intermediate checkpoints, and evaluation code. The release is positioned as a step toward genuine AI sovereignty for Korea's research and industry communities.
Korean AI startup VIDRAFT (CEO Minsik Kim) publicly released its independently developed foundation model, Aether-7B-5Attn, on the global AI platform Hugging Face on July 20, 2026 — marking a significant moment for the Korean AI ecosystem's push toward transparent, reproducible model development.
Aether-7B-5Attn is published under the Apache-2.0 license, but VIDRAFT's ambition goes well beyond a standard weight release. The company has opened up the complete training pipeline: data composition recipes, tokenization procedures, training code, configuration settings, training logs, intermediate checkpoints, and evaluation scripts. Both a base model and an instruction-tuned model, along with a live demo, are freely accessible on Hugging Face.
The model itself is built on a Mixture-of-Experts (MoE) architecture with a total of 6.59 billion parameters, of which approximately 2.98 billion are activated per token — a design choice that boosts inference efficiency without sacrificing overall model capacity.
One of Aether-7B-5Attn's most distinctive technical features is its integration of five heterogeneous attention mechanisms within a single model: Full Attention, Differential Attention, Sliding Window, NSA-family sparse attention, and a Hybrid approach. To prevent any single attention type from dominating a particular depth range, VIDRAFT arranged the model's 49 layers using a 7×7 Latin square structure.
Training was conducted on 16 NVIDIA B200 GPUs over a corpus of approximately 144.2 billion tokens. The data mixture deliberately prioritizes mathematics (37.8%), Korean (21.6%), and English (21.6%), reflecting VIDRAFT's intent to build a bilingual Korean-English foundation model rather than an English-centric one.
Seoul-based VIDRAFT describes itself as a Pre-AGI deep-tech company working across AI and quantum-computing-enabled research in fields including physics, chemistry, life sciences, and pharmaceuticals. The company's broader technical roadmap involves a four-layer integrated AGI system comprising the Darwin model family, the AETHER architecture, the PROMETHEUS world model, and HEPHAESTUS embodied AI.
For years, the AI industry has used "open-source LLM" loosely — often describing models that release weights but withhold training data, code, and methodology. That gap makes independent verification and reproducibility nearly impossible for outside researchers and organizations. Aether-7B-5Attn is designed to close that gap, aligning with the philosophy embodied by models such as Allen AI's OLMo: if you can't reproduce it, you can't truly own it.
This philosophy connects directly to the growing global debate around sovereign AI. VIDRAFT's CEO Minsik Kim argued in the announcement that downloading someone else's weights is not enough to establish AI sovereignty. "Disclosing the data, code, and training process so that anyone can rebuild and verify it is the essence of open source and the starting point of sovereign AI," Kim said, according to ZDNet Korea. He added that he hopes the release will "elevate the conversation around transparency, reproducibility, and independent model-building capacity within Korea's AI ecosystem."
Until now, fully open foundation models of this kind — where the entire manufacturing process is auditable — have been driven primarily by overseas research institutions or national-level consortia such as OLMo, Apertus, and LLM-jp. VIDRAFT's release demonstrates that a single Korean startup can pursue the same standard of openness, offering universities, enterprises, government agencies, and research institutes a foundation they can adapt to their own linguistic, industrial, and security requirements without dependence on foreign Big Tech.
VIDRAFT has previously reported that its Darwin model family has surpassed one million cumulative downloads on Hugging Face, and the company has claimed benchmark achievements including a first-place ranking on the K-AI Leaderboard overall, a 90.9% score on GPQA Diamond, 14 consecutive top rankings on the Polaris global drug-discovery leaderboard, and first place on a metacognition leaderboard — all in 2026. The company is also developing AI infrastructure technologies including the VKAE inference acceleration engine and the VKUE lightweight runtime engine.
Q: What makes Aether-7B-5Attn different from other open-source LLMs?
A: Most models marketed as "open-source" release only weights. Aether-7B-5Attn also publishes training data recipes, full training code, hyperparameters, logs, intermediate checkpoints, and evaluation procedures, enabling complete external reproducibility.
Q: Where can I access VIDRAFT's Aether-7B-5Attn model?
A: The base model, instruction-tuned model, and a live demo are all available on Hugging Face under the Apache-2.0 license, free for anyone to use.
Q: How does this release relate to AI sovereignty?
A: VIDRAFT argues that true AI sovereignty requires the ability to understand and independently reproduce a model's full development process — not just possess its weights. By releasing every component of the training pipeline, the company aims to help Korean institutions build AI suited to their own language, industry, and security needs without relying on foreign platforms.
Source: ZDNet Korea (2026-07-20) — original article