A Korean AI startup just raised the bar for transparency in open-source AI development.
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has publicly released its open-source foundation model Aether-7B-5Attn on Hugging Face under the Apache 2.0 license. Unlike most models that release weights alone, Aether-7B-5Attn ships with training data recipes, full training code, hyperparameters, training logs, intermediate checkpoints, and evaluation code. The release is positioned as a meaningful step toward genuine AI sovereignty for Korea's research and technology ecosystem.
Korean AI startup VIDRAFT has released Aether-7B-5Attn, a fully open-source large language model, on the global AI platform Hugging Face — making it one of the most transparent foundation model releases to come out of a single Korean company to date. The announcement was reported by 전자신문 on July 20, 2026.
VIDRAFT, led by CEO Minsik Kim and headquartered in Seoul, unveiled Aether-7B-5Attn as a fully reproducible foundation model. The release goes well beyond what most companies mean when they say "open source." In the LLM landscape, the majority of so-called open-source models are in practice "open-weight" releases — users can download and run the model, but cannot independently verify or recreate how it was built. Aether-7B-5Attn is designed to close that gap entirely.
The model's public repository includes not just weights, but the complete training data recipe, tokenization pipeline, training code, configuration values, logs, and evaluation procedures. This means any researcher, institution, or company can trace and reproduce the entire process from scratch — a standard more closely aligned with projects like Allen AI's OLMo than with typical commercial releases.
On the architectural side, Aether-7B-5Attn adopts a Mixture-of-Experts (MoE) structure with approximately 6.59 billion total parameters, of which roughly 2.98 billion are activated per token — a design choice aimed at improving inference efficiency without sacrificing model scale. What makes the architecture particularly distinctive is its integration of five different attention mechanisms within a single model: Full Attention, Differential Attention, Sliding Window, NSA-family sparse attention, and a Hybrid approach. The model's 49 layers are arranged in a 7×7 Latin Square pattern, ensuring no single attention type dominates any particular depth of the network.
Training was conducted on 16 NVIDIA B200 GPUs across approximately 144.2 billion tokens. The training data composition reflects VIDRAFT's intention to build a bilingual foundation model rather than an English-centric one: mathematics accounts for roughly 37.8% of the dataset, with Korean and English each comprising approximately 21.6%. Both a base model and an instruct model are available on Hugging Face, along with a live demo.
The model is released under the Apache 2.0 license, making it freely available for commercial and research use.
The release arrives at a moment when the concept of "sovereign AI" is gaining traction globally — and VIDRAFT's framing of this launch speaks directly to that conversation. AI sovereignty, in meaningful terms, is not simply about a country or company possessing model weights. It requires the capacity to understand, audit, and independently reproduce the underlying data, code, and training process.
Until now, fully open foundation models of this nature have largely come from overseas research institutions or national-level consortia — projects like OLMo (Allen AI), Apertus, and LLM-jp. VIDRAFT's Aether-7B-5Attn represents a rare instance of a single startup pursuing the same level of openness and reproducibility, signaling that Korean AI development is ready to compete on transparency as well as performance.
CEO Minsik Kim stated that simply downloading someone else's weights is not sufficient to achieve AI sovereignty. He emphasized that genuine openness — making data, code, and the training process fully accessible so that anyone can rebuild and verify the model — is both the essence of open source and the starting point for sovereign AI. He expressed hope that the release would elevate discussions around transparency, reproducibility, and independent model development within Korea's AI ecosystem.
VIDRAFT itself operates across a broader technology stack. The company's publicly stated roadmap includes the Darwin model family, the AETHER architecture, the PROMETHEUS world model, and the HEPHAESTUS embodied AI system — described as a four-layer integrated AGI framework. The company has previously reported that the Darwin model family has surpassed one million cumulative downloads on Hugging Face, and has disclosed benchmark results including a top ranking on the K-AI leaderboard and a 90.9% score on GPQA Diamond.
Q: What makes Aether-7B-5Attn different from other open-source LLMs?
A: Most LLMs described as "open source" only release model weights. Aether-7B-5Attn also releases training data recipes, training code, hyperparameters, logs, intermediate checkpoints, and evaluation code, making the full training process reproducible by external parties.
Q: What attention mechanisms does Aether-7B-5Attn use?
A: The model integrates five types of attention mechanisms — Full Attention, Differential Attention, Sliding Window, NSA-family sparse attention, and a Hybrid method — within a single architecture, arranged across 49 layers using a 7×7 Latin Square structure.
Q: Where can I access the Aether-7B-5Attn model?
A: Both the base model and instruct model, along with a live demo, are publicly available on Hugging Face under the Apache 2.0 license.
Source: 전자신문 (2026-07-20) — original article