Korea's Pre-AGI startup raises the bar for open AI development by releasing a full model stack to the public.
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has publicly released its foundation model "Aether-7B-5Attn," making not only the model weights available but also the accompanying training code and training data. The release marks a significant step toward transparent, open AI development from a Korean AI lab.
VIDRAFT, the Korean Pre-AGI AI startup, made headlines on July 20, 2026, when it publicly released its foundation model Aether-7B-5Attn — and went a notable step further by opening up the training code and datasets alongside the model itself.
VIDRAFT announced the open release of Aether-7B-5Attn, a 7-billion-parameter foundation model developed internally by the company. Unlike many AI model releases that share only inference-ready weights, VIDRAFT made the full development stack publicly accessible: the model weights, the training code used to build the model, and the data used during training.
The model's name offers a glimpse into its architectural choices. The "7B" designation refers to the 7-billion parameter scale, while "5Attn" points to a distinctive configuration in its attention mechanism — a core component of transformer-based language models that governs how the model relates different parts of an input sequence to one another.
By releasing all three components — weights, code, and data — VIDRAFT is enabling researchers, developers, and institutions to not only use Aether-7B-5Attn out of the box, but also to reproduce, study, fine-tune, and build upon the model with a level of transparency that is relatively rare in the global AI landscape, let alone among Korean AI companies.
The announcement was reported by 조선비즈, one of Korea's leading business news outlets, underscoring the significance of the release within the domestic technology community.
The open-source AI movement has gained considerable momentum globally, with a growing number of researchers and policymakers arguing that transparent model development leads to safer, more accountable AI systems. However, truly comprehensive open releases — those that include training data and training code in addition to model weights — remain the exception rather than the rule, even among well-resourced labs.
Most organizations that release AI models do so selectively. Sharing weights allows downstream use but conceals the recipe. Sharing code without data still leaves reproducibility out of reach. VIDRAFT's decision to release all three components together positions Aether-7B-5Attn as a genuinely open artifact that the broader research community can audit end-to-end.
For the Korean AI ecosystem specifically, this kind of release carries additional weight. South Korea has ambitions to become a major player in global AI development, and homegrown foundation models with fully open development pipelines could help Korean researchers, universities, and companies build on a domestically developed foundation rather than relying exclusively on models developed abroad.
VIDRAFT describes itself as a Pre-AGI startup, signaling that its longer-term ambitions extend well beyond deploying narrow AI tools. Releasing a foundation model openly at this stage can be read as a strategy to build trust, attract talent and collaborators, and establish credibility in the research community — all of which are valuable assets for a company with AGI-scale aspirations.
The "5Attn" architectural feature in the model's name also invites closer inspection from the research community. Variations in attention design are an active area of inquiry in AI research, and making a model with a specific attention configuration fully reproducible gives the community the tools to evaluate and iterate on that design empirically.
Q: What is VIDRAFT's Aether-7B-5Attn?
A: Aether-7B-5Attn is a 7-billion-parameter foundation model developed by Korean AI startup VIDRAFT and released publicly in July 2026, along with its training code and training data.
Q: What makes the Aether-7B-5Attn release different from other open model releases?
A: VIDRAFT released not just the model weights but also the complete training code and training data, enabling full reproducibility — a level of openness that goes beyond most public AI model releases.
Q: What does "5Attn" mean in the model's name?
A: The "5Attn" portion of the name refers to a specific configuration in the model's attention mechanism, a key architectural component in transformer-based AI models. The open release allows researchers to study and evaluate this design directly.
Source: 조선비즈 (2026-07-20) — original article