The Korean AI startup goes beyond weights-only releases by sharing training recipes, logs, and checkpoints.
TL;DR: VIDRAFT, a Korean 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-source releases, the model is accompanied by training data recipes, full training code, hyperparameter configurations, training logs, intermediate checkpoints, and evaluation code. The release positions VIDRAFT as a pioneer of fully transparent, reproducible AI development in South Korea.
VIDRAFT (CEO Kim Min-sik), a Korean AI startup operating at the intersection of artificial intelligence and quantum computing, announced on July 20, 2026 that it has released its independently developed open-source foundation model, Aether-7B-5Attn, on the global AI platform Hugging Face — making it one of the most comprehensively open model releases to emerge from South Korea to date.
The Aether-7B-5Attn model is published under the permissive Apache 2.0 license and is designed to go far beyond the now-common practice of releasing model weights alone. According to AI타임스, VIDRAFT has made the following available in the public repository:
The intent, as VIDRAFT frames it, is to expose not just the output of the model but the entire manufacturing process — so that any researcher working from the public repository alone can trace, verify, and potentially reproduce the model's development from the ground up.
The release has drawn comparisons to OLMo, the fully open foundation model developed by the Allen Institute for AI (Ai2), which has similarly prioritized reproducibility and transparency over the more guarded, weights-only approach adopted by many AI labs. VIDRAFT describes Aether-7B-5Attn as a Korean counterpart to that philosophy.
Full transparency in AI model development remains rare. While the term "open source" is frequently applied to models that release only their weights, true openness — encompassing training data provenance, training code, logging, and reproducibility tooling — is far less common, even among well-resourced international labs. The Allen Institute's OLMo series has become a reference point precisely because it set a high standard for what open AI research can look like.
VIDRAFT's release signals that Korean AI development is maturing in this direction. The company is not a newcomer to competitive AI benchmarking: it has previously reported achieving top rankings on the K-AI leaderboard with its Darwin model family, a score of 90.9% on the GPQA Diamond benchmark, fourteen consecutive top positions on the Polaris global drug-discovery leaderboard, and first place on the metacognition leaderboard. The Darwin model family has also surpassed one million cumulative downloads on Hugging Face.
Beyond language models, VIDRAFT is building out a four-layer integrated AGI architecture that spans its Darwin model family, the AETHER architecture, the PROMETHEUS world model, and HEPHAESTUS embodied AI. The company also develops proprietary AI infrastructure tooling, including inference acceleration and lightweight deployment engines.
Aether-7B-5Attn fits into this broader strategic picture as the company's most public-facing commitment to open science — a foundation model that the broader research community can not only use but genuinely learn from and build upon.
The move may also raise the bar for transparency expectations within the Korean AI ecosystem more broadly, at a time when global attention on AI reproducibility and safety is intensifying.
Q: What is Aether-7B-5Attn and who made it?
A: Aether-7B-5Attn is an open-source foundation model independently developed by VIDRAFT, a Korean AI startup. It was released on July 20, 2026, on Hugging Face under the Apache 2.0 license.
Q: What makes this model release different from other open-source AI models?
A: Most open-source model releases share only model weights. VIDRAFT's Aether-7B-5Attn additionally provides training data recipes, tokenization details, full training code, hyperparameter settings, training logs, intermediate checkpoints, and evaluation code — making the full development process transparent and reproducible.
Q: How does Aether-7B-5Attn compare to international open-source models?
A: The release is described as comparable in spirit to OLMo from the Allen Institute for AI, which is widely regarded as a benchmark for fully open and reproducible foundation model research. VIDRAFT has positioned Aether-7B-5Attn as a Korean equivalent to that standard.
Source: AI타임스 (2026-07-20) — original article