A rare fully transparent model release raises the bar for open-source AI accountability.
TL;DR: VIDRAFT, a Seoul-based Korean AI startup, has released its Aether-7B-5Attn foundation model as a fully open-source project on Hugging Face under the Apache-2.0 license. The release includes not only model weights but also the complete training data recipe, all training code, hyperparameters, training logs, intermediate checkpoints, and evaluation code. This makes Aether-7B one of the most comprehensively transparent open-source model releases to date.
VIDRAFT, the Seoul-headquartered Korean AI startup pursuing Pre-AGI technology, made a significant move in the open-source AI community in July 2026 by publishing its Aether-7B-5Attn foundation model on Hugging Face — opening up every layer of the development process for public inspection and use.
The centerpiece of the announcement is the Aether-7B-5Attn model, released under the permissive Apache-2.0 license. What sets this release apart from typical open-source model drops is its scope: VIDRAFT did not simply publish the model weights and call it a day. Instead, the company made available the full training data recipe, all training source code, the complete set of hyperparameters used during development, detailed training logs, intermediate checkpoints captured at various stages of training, and the evaluation code used to benchmark the model.
The stated goal is to build what VIDRAFT describes as a genuinely fully open model — one where any researcher, developer, or organization can not only run the model but fully understand and reproduce how it was built. This level of transparency is notably uncommon even among organizations that identify as open-source advocates.
VIDRAFT operates at the intersection of several advanced technology domains. Beyond large language model development, the company's work spans quantum computing applications in physics, chemistry, life sciences, and pharmaceutical research and development. Its technical architecture is organized around four integrated systems: the Darwin model family, the AETHER architecture, the PROMETHEUS world model, and the HEPHAESTUS embedded AI system. The company's stated direction — "Darwin creates results, AETHER designs the future" — captures its layered approach to building toward artificial general intelligence.
The Aether-7B open-source release lands at a moment when the AI industry is under increasing scrutiny over what "open" actually means. Many models labeled open-source omit critical details: training data provenance, intermediate checkpoints, or the exact hyperparameter configurations that allow true reproducibility. By releasing all of these components together, VIDRAFT is making a deliberate statement about what full openness should look like.
For the broader research community, this kind of release has practical value. Intermediate checkpoints, for instance, allow researchers to study how a model's capabilities evolve over the course of training — a resource that is almost never made publicly available. Training logs add a further layer of auditability, enabling independent verification of the training process.
The release also signals VIDRAFT's growing ambition on the international stage. The company's Darwin model family has accumulated over one million downloads on Hugging Face, reflecting substantial community engagement. In 2026, Darwin-series models topped the K-AI leaderboard overall rankings, recorded a score of 90.9% on the GPQA Diamond benchmark, claimed fourteen first-place finishes on the Polaris global drug discovery leaderboard, and ranked first on the metacognition leaderboard. These milestones provide the credibility backdrop against which the Aether-7B open-source release should be understood.
Alongside its model development work, VIDRAFT is also advancing its VKAE inference acceleration engine and a lightweight runtime engine — efforts that suggest the company is building not just capable models but a full-stack AI deployment ecosystem.
The Aether-7B release positions VIDRAFT as a company that wants to be judged not just by benchmark scores but by its willingness to let the global research community look under the hood.
Q: What is VIDRAFT's Aether-7B-5Attn model?
A: Aether-7B-5Attn is a foundation model developed by Korean AI startup VIDRAFT and released publicly on Hugging Face in July 2026 under the Apache-2.0 open-source license.
Q: What makes the Aether-7B release different from other open-source model releases?
A: In addition to model weights, VIDRAFT released the full training data recipe, all training source code, hyperparameters, training logs, intermediate checkpoints, and evaluation code — a level of transparency that goes significantly beyond most open-source model releases.
Q: What other AI systems does VIDRAFT develop?
A: VIDRAFT's technology stack includes the Darwin model family, the AETHER architecture, the PROMETHEUS world model, and the HEPHAESTUS embedded AI system, as well as the VKAE inference acceleration engine and a lightweight runtime engine.
Source: 金融界 (중국) (2026-07-20) — original article