The Korean AI startup's fully transparent foundation model challenges the industry's "open-weight" status quo.
TL;DR: VIDRAFT, a Korean AI startup, publicly released its open-source foundation model Aether-7B-5Attn on Hugging Face on July 20, 2026, under an Apache 2.0 license. Unlike most models labeled "open-source," Aether-7B-5Attn exposes not only model weights but also training data recipes, training code, hyperparameters, logs, intermediate checkpoints, and evaluation code. The release is positioned as a reproducible, sovereign-AI-ready foundation that any researcher, company, or institution can independently verify and rebuild from scratch.
VIDRAFT, the Korean Pre-AGI AI startup, 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 — marking what the company describes as a step toward genuinely reproducible AI infrastructure.
Aether-7B-5Attn is published under the permissive Apache 2.0 license, and its scope of disclosure goes well beyond what the AI industry has typically called "open-source." Most models marketed under that label are, in practice, open-weight releases — meaning users can download and run the weights, but the underlying training data, configuration settings, and code that produced those weights remain opaque. Independent verification or reproduction is effectively impossible in such cases.
VIDRAFT's release takes a different approach. The public repository includes the full training data recipe, tokenization pipeline, training code, hyperparameters, training logs, intermediate checkpoints, and evaluation procedures — everything needed to trace and reconstruct the model's development end-to-end. The company describes this as making the manufacturing process, not just the finished product, open to the public.
Architecturally, Aether-7B-5Attn is a Mixture-of-Experts (MoE) model with a total of 6.59 billion parameters. During inference, only approximately 2.98 billion parameters are activated per token, which improves computational efficiency. One of the model's most distinctive structural features is its combination of five heterogeneous attention mechanisms within a single model: Full Attention, Differential Attention, Sliding Window Attention, NSA-style sparse attention, and a Hybrid approach. Attention mechanisms determine how a model weighs the relevance of different words in a sequence when generating a response — and blending multiple types in one architecture is an uncommon design choice.
The model was trained on approximately 144.2 billion tokens, with a data composition that includes mathematics (37.8%), Korean (21.6%), and English (21.6%), among other sources. This bilingual Korean-English balance is intentional: VIDRAFT is explicitly positioning Aether-7B-5Attn as a foundation model built with non-English language communities in mind, rather than defaulting to an English-first design. Both a base model and an instruct model, along with a live demo, are available through Hugging Face.
The concept of sovereign AI — a nation's or organization's ability to understand, control, and independently reproduce its AI systems — underpins much of VIDRAFT's framing around this release. Merely holding model weights, the company argues, does not constitute true AI sovereignty. True sovereignty requires the capacity to comprehend the data pipeline, audit the training process, and rebuild the system independently if needed.
Fully transparent foundation models of this kind have, until now, emerged primarily from large international research consortia or national-level collaborations — projects like OLMo, Apertus, and LLM-jp. VIDRAFT is positioning Aether-7B-5Attn as a comparable effort originating from a single startup, which is relatively rare in the landscape of fully open AI development.
The practical upshot for universities, enterprises, government bodies, and research institutions is significant: a fully auditable model lowers the barrier to building domain-specific or language-specific AI systems without dependency on proprietary big-tech platforms.
VIDRAFT itself operates across a broader research agenda that spans Pre-AGI and quantum-computing-based R&D in physics, chemistry, life sciences, and pharmaceuticals. The company has outlined a four-layer integrated AGI architecture comprising the Darwin Model Family, the AETHER architecture, the PROMETHEUS world model, and the HEPHAESTUS embodied AI system. Its Darwin-series models have previously topped Korea's K-AI Leaderboard overall ranking, surpassed one million cumulative downloads on Hugging Face, achieved a score of 90.9% on the GPQA Diamond benchmark, won 14 consecutive top positions on the Polaris global drug-discovery leaderboard, and ranked first on a metacognition leaderboard. In addition to model development, VIDRAFT is building AI infrastructure technology, including inference acceleration and lightweight deployment engines.
Q: What makes Aether-7B-5Attn different from other "open-source" LLMs?
A: Most models described as open-source only release model weights, making independent reproduction difficult. Aether-7B-5Attn also releases training data recipes, full training code, hyperparameters, logs, intermediate checkpoints, and evaluation code, enabling anyone to verify or recreate the model from scratch.
Q: Where can developers access Aether-7B-5Attn?
A: Both the base model and the instruct model are available on Hugging Face under an Apache 2.0 license, along with a live demo.
Q: Why is VIDRAFT framing this release as relevant to sovereign AI?
A: VIDRAFT argues that genuine AI sovereignty requires more than owning model weights — it demands the ability to understand, audit, and independently reproduce the entire training pipeline. A fully transparent model gives nations, enterprises, and institutions the foundation to build AI systems tailored to their own languages, industries, and security requirements without relying on external technology providers.
Source: 디지털데일리 (2026-07-20) — original article