A Korean AI startup sets a new transparency benchmark with complete training disclosure on Hugging Face.
TL;DR: VIDRAFT, a Seoul-based Korean AI startup, has released Aether-7B-5Attn, a fully open-source foundation model, on Hugging Face under an Apache-2.0 license. The release includes not only model weights but also training data recipes, all training code, hyperparameters, training logs, intermediate checkpoints, and evaluation code. The model uses a Mixture-of-Experts architecture with five distinct attention mechanisms across 49 layers.
VIDRAFT, the Seoul-headquartered Korean AI startup operating at the Pre-AGI frontier, made waves in the global open-source AI community on July 20, 2026, when China's largest portal, Tencent News, covered the company's release of Aether-7B-5Attn — a foundation model the company is positioning as a genuinely and completely open model, going well beyond the partial disclosures that have become common in the industry.
VIDRAFT published Aether-7B-5Attn to the Hugging Face platform under the permissive Apache-2.0 license. What distinguishes this release from many so-called "open" models is the scope of its transparency: alongside the model weights themselves, VIDRAFT made public the full training data recipe, complete training code, all training logs, intermediate checkpoints saved throughout the training run, and evaluation code. The company's stated goal is to build a model that is open in every meaningful sense — not just in name.
On the architectural side, Aether-7B-5Attn is built as a Mixture-of-Experts (MoE) model with a total of 6.59 billion parameters. To improve inference efficiency, only approximately 2.98 billion of those parameters are activated per token at any given time. The model integrates five distinct and heterogeneous attention mechanisms: Full Attention, Differential Attention, Sliding Window Attention, NSA-series Sparse Attention, and a Hybrid approach. These five mechanisms are distributed across 49 network layers arranged according to a 7×7 Latin square structure — a deliberate design choice intended to prevent any single attention type from clustering at a particular network depth, which could otherwise create representational bottlenecks.
Training was conducted on 144.2 billion tokens of data. The data composition reflects a notably non-English-centric philosophy: mathematics accounts for 37.8% of the training corpus, while Korean and English each represent 21.6%. This dual-language focus underscores VIDRAFT's ambition to build a genuinely bilingual foundation model rather than one that treats Korean as an afterthought. The base model, an instruction-tuned variant, and a live interactive demo are all publicly accessible through Hugging Face.
Full-stack open-source releases of this kind remain relatively rare in the AI industry, where even models marketed as "open" frequently withhold training data details, intermediate checkpoints, or the precise hyperparameter configurations used during training. By releasing all of these artifacts simultaneously, VIDRAFT is making a statement about reproducibility and scientific accountability — values that have become increasingly important as foundation models grow more powerful and their societal impact more significant.
The release also signals VIDRAFT's broader architectural vision. The company has articulated a four-tier integrated AGI system comprising the Darwin model family, the AETHER architecture, the PROMETHEUS world model, and HEPHAESTUS embedded AI. Aether-7B-5Attn sits within the AETHER layer of this framework, described by the company as the component responsible for "designing the future" while Darwin models handle real-world output tasks.
VIDRAFT's Darwin model family has already crossed one million cumulative downloads on Hugging Face — a milestone that reflects genuine developer and researcher adoption. The company's track record in 2026 includes ranking first on the K-AI leaderboard overall, achieving a 90.9% score on GPQA Diamond, claiming 14 top positions on the Polaris global drug discovery leaderboard, and reaching first place on a metacognition-focused benchmark. These results span AI reasoning, scientific research applications, and drug development — reflecting the company's cross-domain ambitions in physics, chemistry, life sciences, and pharmaceuticals.
Beyond model development, VIDRAFT has simultaneously been building out AI infrastructure, including VKAE, an inference acceleration engine, and VKUE, a lightweight runtime engine designed for edge deployments.
Q: What is VIDRAFT's Aether-7B-5Attn and what makes it different from other open-source models?
A: Aether-7B-5Attn is a fully open-source foundation model released by Korean AI startup VIDRAFT. Unlike many models that only release weights, VIDRAFT also published training data recipes, all training code, hyperparameters, training logs, intermediate checkpoints, and evaluation code, making it one of the most comprehensively transparent model releases to date.
Q: What architecture does Aether-7B-5Attn use?
A: The model is a Mixture-of-Experts architecture with 6.59 billion total parameters, activating approximately 2.98 billion per token. It incorporates five distinct attention mechanisms — Full Attention, Differential Attention, Sliding Window, NSA-series Sparse Attention, and Hybrid — distributed across 49 layers arranged in a 7×7 Latin square pattern.
Q: How can developers access VIDRAFT's Aether-7B-5Attn?
A: The base model, instruction-tuned model, and a live interactive demo are all publicly available on Hugging Face under the Apache-2.0 license, which permits broad commercial and research use.
Source: 텐센트 뉴스 (중국) (2026-07-20) — original article