Korea's deep-tech startup bets on full-stack transparency to advance the open-source AI ecosystem.
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. The 6.59-billion-parameter mixture-of-experts model ships alongside training data composition details, training code, hyperparameters, logs, and evaluation code. Both base and instruct variants, plus a live demo, are available on Hugging Face now.
Korean AI startup VIDRAFT (CEO Kim Min-sik) launched its latest open-source foundation model, Aether-7B-5Attn, on the global AI platform Hugging Face on July 20, 2026 — marking one of the most transparent model releases to come out of the Korean AI scene to date. Unlike releases that share only model weights, VIDRAFT has made the full development record publicly accessible under the permissive Apache-2.0 license.
VIDRAFT's new model, Aether-7B-5Attn, is built on a Mixture-of-Experts (MoE) architecture with a total of 6.59 billion parameters. During inference, approximately 2.98 billion parameters are activated per token — a design choice intended to maximize computational efficiency without sacrificing model capacity.
One of the model's defining architectural features is its combination of five distinct attention mechanisms: Full, Differential, Sliding Window, NSA-family sparse, and Hybrid attention. These are distributed across 49 layers arranged in a 7×7 Latin square pattern, ensuring that each attention type operates evenly across different depths of the network rather than clustering in specific regions.
The training corpus spanned roughly 144.2 billion tokens, with the dataset weighted toward mathematics (37.8%), Korean (21.6%), and English (21.6%). This bilingual emphasis means the model is explicitly designed to perform well in both Korean and English-language environments — a practical consideration for developers targeting either market.
What sets this release apart from a typical open-weight drop is the accompanying documentation. VIDRAFT has published the training data composition methodology, the actual training code, hyperparameters used during the training run, training logs, and evaluation code. The company says this level of openness is intended to allow external researchers and organizations to independently verify the model's technical structure and reproduce research built on it.
Transparency in AI development has become an increasingly urgent conversation globally, and VIDRAFT's approach with Aether-7B-5Attn directly addresses one of the field's core tensions: the gap between "open-source" models that release only weights and genuinely reproducible AI research.
By releasing the full development pipeline — not just the end artifact — VIDRAFT is positioning Aether-7B-5Attn as a reference point for what technical transparency can look like from a single startup. For research institutions and enterprises wary of dependency on opaque external AI systems, a fully documented training process offers a meaningful alternative. Organizations can audit the model's origins, adapt training procedures to their own data requirements, and build internal AI capabilities with a reduced reliance on black-box solutions.
The release also carries significance for AI security. When a model's training lineage is fully traceable, it becomes considerably easier for adopters to assess risk, verify compliance, and customize behavior for sensitive applications.
VIDRAFT describes itself as a deep-tech company operating at the intersection of quantum computing research, AI model development, and R&D solutions for physics, chemistry, and life sciences. The company has previously introduced the Darwin model family and its own system architectures, achieving results on major AI benchmarks and drug-discovery-related evaluations. Its Hugging Face presence has accumulated over one million cumulative downloads. Beyond model development, the company is also pursuing AI infrastructure research, including its inference acceleration engine VKAE and lightweight runtime engine VKUE.
CEO Kim Min-sik commented that being able to transparently verify not just model weights but also data, code, and the full training process forms the foundation of a healthy open-source ecosystem. He expressed hope that the release will spark broader discussion within Korea's AI research community around technical transparency and independent model development capability.
Q: What is VIDRAFT's Aether-7B-5Attn model?
A: Aether-7B-5Attn is a 6.59-billion-parameter open-source foundation model released by Korean AI startup VIDRAFT. It uses a Mixture-of-Experts architecture combined with five attention mechanisms and is available on Hugging Face under the Apache-2.0 license.
Q: What makes this release different from other open-source model releases?
A: In addition to model weights, VIDRAFT published the training data composition methodology, training code, hyperparameters, training logs, and evaluation code — enabling external researchers and organizations to independently verify and reproduce the model's development process.
Q: Who is VIDRAFT, and what else has the company built?
A: VIDRAFT is a Korean deep-tech startup led by CEO Kim Min-sik, focused on AI model development, quantum computing research, and R&D solutions for the physical and life sciences. The company has previously released the Darwin model family, developed AI infrastructure tools including the VKAE and VKUE engines, and surpassed one million cumulative downloads on Hugging Face.
Source: 동아일보 (2026-07-20) — original article