A Korean AI startup goes beyond open weights — releasing training data, code, and logs for full reproducibility.
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 open-weight models, VIDRAFT's release includes training data recipes, full training code, hyperparameters, training logs, intermediate checkpoints, and evaluation code. The company describes the release as a concrete step toward "sovereign AI" — where anyone can rebuild and verify a model from scratch.
Korean AI startup VIDRAFT made a significant move in the open-source AI landscape on July 20, 2026, releasing its independently developed foundation model "Aether-7B-5Attn" to the public via Hugging Face. The release goes well beyond the industry norm of sharing only model weights, positioning VIDRAFT as one of the few organizations globally — and the first in Korea — to pursue a fully transparent, fully reproducible open-source foundation model.
VIDRAFT's Aether-7B-5Attn is published under the Apache-2.0 license, meaning anyone can freely use, modify, and distribute it. But the license is arguably the least distinctive part of this release. What sets it apart is the scope of what VIDRAFT made available alongside the model weights: the training data recipe, complete training code, all configuration values, training logs, intermediate checkpoints, and evaluation code.
According to the company, the vast majority of models commonly described as "open source" are in fact "open weight" — they share the final trained model but reveal nothing about how it was built. VIDRAFT's approach draws a deliberate contrast with that norm. The company compares its methodology to that of Allen AI's OLMo project, one of the few internationally recognized examples of a truly open foundation model, alongside Apertus and LLM-jp.
On the technical side, Aether-7B-5Attn uses a Mixture-of-Experts (MoE) architecture with approximately 6.59 billion parameters and integrates five distinct attention mechanisms. The model was trained on roughly 144.2 billion tokens. The training data composition reflects a multilingual, math-heavy approach: mathematics accounts for 37.8% of the data, while Korean and English each make up 21.6%.
VIDRAFT CEO Kim Min-sik stated: "Simply downloading someone else's weights does not complete AI sovereignty. Making data, code, and the training process open so that anyone can rebuild and verify it — that is the essence of open source and the starting point of sovereign AI."
The concept of "sovereign AI" has gained traction globally as governments and organizations grow wary of depending on foreign-developed, closed AI systems. True sovereignty, in this framing, requires not just access to a working model but the ability to understand, audit, reproduce, and modify it independently.
Most large language models released to date — even those labeled open source — fall short of that bar. When only the final weights are shared, the community must take the developer's word on training choices, data quality, and safety measures. Full reproducibility, by contrast, allows independent researchers, companies, and governments to verify claims, identify biases, and adapt the model to local needs without black-box dependencies.
VIDRAFT's release is notable for occurring within Korea, where the domestic AI sector has primarily relied on adapting foreign foundation models rather than building and fully disclosing homegrown ones. By publishing a verifiable training pipeline alongside the model itself, VIDRAFT is making a public case that Korean organizations need not treat foundation-model development as someone else's domain.
The fully open foundation model space has until now been dominated by academic consortia and research institutes outside Korea — Allen AI, the Apertus project, and Japan's LLM-jp among them. VIDRAFT's entry into this category represents a broadening of that ecosystem.
Q: What is VIDRAFT's Aether-7B-5Attn, and where can I access it?
A: Aether-7B-5Attn is a fully open-source foundation model developed by Korean AI startup VIDRAFT. It is available on Hugging Face under the Apache-2.0 license, with training data recipes, code, hyperparameters, logs, checkpoints, and evaluation code all included.
Q: How does VIDRAFT's model release differ from typical open-source LLMs?
A: Most models described as open source only release trained weights, a practice known as "open weight." VIDRAFT's release additionally provides the full training data recipe, code, configuration, logs, and evaluation pipeline, enabling independent verification and reproduction of the model from scratch.
Q: What does VIDRAFT mean by "sovereign AI"?
A: According to VIDRAFT CEO Kim Min-sik, AI sovereignty cannot be achieved simply by downloading another organization's model weights. True sovereignty requires that data, code, and training processes be fully transparent so that any party can independently rebuild and verify the model — a standard VIDRAFT aims to meet with this release.
Source: 중앙일보 (2026-07-20) — original article