The startup goes beyond sharing model weights — publishing data pipelines, training code, and checkpoints too.
TL;DR: VIDRAFT, a Korean AI startup, has released its open-source foundation model Aether-7B-5Attn on Hugging Face under the Apache-2.0 license. Unlike most open-source LLMs that share only model weights, VIDRAFT has also made public its training data composition, training code, hyperparameters, training logs, intermediate checkpoints, and evaluation code. CEO Minsik Kim stated that true AI sovereignty begins with the ability to understand and reproduce the full training process, not just possess model weights.
Korean AI startup VIDRAFT publicly launched its homegrown open-source foundation model, Aether-7B-5Attn, on the global AI platform Hugging Face on July 20, 2026 — marking what the company describes as a significant step toward reproducible, transparent AI development in South Korea.
VIDRAFT released Aether-7B-5Attn under the permissive Apache-2.0 license, making the model freely available for both research and commercial use. What sets this release apart from typical open-source AI drops is the sheer depth of what has been shared.
Most open-source large language models (LLMs) release only their trained weights — the final output of the training process — leaving researchers and developers unable to independently verify how a model was built or replicate its development from scratch. VIDRAFT's release breaks from that convention by also providing the training data composition methodology, training code, hyperparameters, training logs, intermediate checkpoints, and evaluation code. Together, these components allow anyone to reconstruct and verify the model's development pipeline end to end.
Architecturally, Aether-7B-5Attn uses a Mixture-of-Experts (MoE) structure and combines five distinct attention mechanisms to improve inference efficiency. The model was trained on approximately 144.2 billion tokens, with the dataset spanning mathematics, Korean, and English, among other domains — reflecting an intentional focus on multilingual and technical capability.
VIDRAFT CEO Minsik Kim explained the company's rationale plainly: "AI sovereignty starts not just from securing model weights, but from a foundation where you can understand, and reproduce, both the data and the training process." He added that he hopes this release will broaden the conversation around transparency and reproducibility within South Korea's AI ecosystem.
The distinction VIDRAFT is drawing — between releasing weights and releasing the full training process — gets at a growing tension in the global AI community. Many models marketed as "open source" offer limited real openness: you can run the model, but you cannot meaningfully audit it, adapt it at a foundational level, or verify the claims its creators make about how it was trained. This has led critics to argue that weight-only releases are closer to open-access than true open-source.
By contrast, VIDRAFT's Aether-7B-5Attn release is designed so that researchers and companies can validate the model in their own environments using the same conditions under which it was originally built. This level of transparency is relatively rare even among globally prominent AI labs, making it a notable move for a startup operating out of South Korea.
For Korea's domestic AI ecosystem specifically, the release carries additional weight. South Korea has significant ambitions in AI development but has historically depended on architectures and training frameworks pioneered abroad. A fully transparent, reproducible Korean-developed foundation model — complete with multilingual Korean-English-math training data — provides local researchers, universities, and enterprises with a genuine starting point for independent model development and experimentation, without being locked into opaque external systems.
VIDRAFT frames this as a matter of AI sovereignty: the ability of a country or organization to not merely use AI, but to understand, verify, and build upon the systems it depends on. In that framing, transparency is not just a technical choice — it is a strategic one.
Q: What is VIDRAFT's Aether-7B-5Attn model?
A: Aether-7B-5Attn is an open-source foundation model developed by Korean AI startup VIDRAFT. It uses a Mixture-of-Experts architecture with five attention mechanisms and was trained on roughly 144.2 billion tokens covering Korean, English, and mathematics.
Q: How is this release different from other open-source LLM releases?
A: Most open-source LLMs share only trained model weights, but VIDRAFT's release also includes training data composition methodology, source code, hyperparameters, training logs, intermediate checkpoints, and evaluation code — enabling independent reproduction and verification of the entire training process.
Q: Where can researchers access Aether-7B-5Attn?
A: The model is available on Hugging Face under the Apache-2.0 license, which permits both research and commercial use.
Source: 한국일보 (2026-07-20) — original article