A Korean AI startup sets a new bar for reproducibility with a completely open MoE foundation model release.
TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has released Aether-7B-5Attn on Hugging Face under an Apache-2.0 license, publishing not just model weights but also training code, data recipes, hyperparameters, full training logs, and intermediate checkpoints. The model is a Mixture-of-Experts architecture with approximately 6.59 billion parameters trained on roughly 144.2 billion tokens. The release is designed to be fully reproducible by the broader research and developer community.
Korean AI startup VIDRAFT has made its Aether-7B-5Attn foundation model fully available to the public on Hugging Face, releasing every artifact needed to reproduce the work — weights, training code, data recipes, hyperparameters, logs, and checkpoints — all under the permissive Apache-2.0 license. The announcement was first covered by English-language tech outlet ReasonCore on July 20, 2026.
The centerpiece of the release is Aether-7B-5Attn, a Mixture-of-Experts (MoE) language model with approximately 6.59 billion parameters. The model's architecture is built around a distinctive 7×7 Latin-square layout that distributes five different attention types across 49 layers — an approach that distinguishes it structurally from more conventional transformer designs. The model was trained on approximately 144.2 billion tokens.
What makes this release particularly notable is its scope. Rather than publishing weights alone — already common practice among open-model developers — VIDRAFT has accompanied Aether-7B-5Attn with the full training codebase, a detailed data recipe, all hyperparameters used during training, complete training logs, and evaluation code. The project also includes intermediate checkpoints saved at steps 110k, 115k, and 162k, with each checkpoint weighing approximately 13 GB.
All of these materials are available under the Apache-2.0 license, one of the most permissive open-source licenses available, allowing commercial use, modification, and redistribution with minimal restrictions.
ReasonCore notes that the current release does not include benchmark scores, and that this build does not implement KV-cache.
Truly reproducible AI model releases remain rare. Many organizations that describe their models as "open" publish only weights or inference code, leaving the research community unable to verify training decisions, audit data choices, or build meaningfully on the original work. VIDRAFT's decision to release the full training stack — including intermediate checkpoints and logs — goes considerably further than most comparable announcements.
The heterogeneous-attention architecture at the heart of Aether-7B-5Attn is also worth attention. The 7×7 Latin-square layout, which systematically varies attention mechanisms across the model's 49 layers, represents a structural design choice that researchers and engineers can now study, replicate, and extend directly from the published code. Without training logs and checkpoints, such architectural innovations are difficult to evaluate properly; with them, the community can trace exactly how the model evolved during training.
For a startup still in the Pre-AGI research phase, this level of openness signals a deliberate strategy: building credibility and community trust by giving researchers and developers every tool they need to engage with the work on its own terms. In a market where large labs often cite safety or competitive concerns to justify partial releases, a fully documented Apache-2.0 drop from a Korean AI startup stands out.
The Apache-2.0 license choice is also significant for commercial developers. Unlike more restrictive open licenses that prohibit commercial use or impose copyleft conditions, Apache-2.0 lets companies integrate, fine-tune, or build products on top of Aether-7B-5Attn without needing to negotiate separate agreements.
Q: What is Aether-7B-5Attn and who made it?
A: Aether-7B-5Attn is a Mixture-of-Experts foundation model with approximately 6.59 billion parameters developed by VIDRAFT, a Korean Pre-AGI AI startup. It features a 7×7 Latin-square layout of five attention types across 49 layers and was trained on roughly 144.2 billion tokens.
Q: What exactly is included in VIDRAFT's open-source release?
A: The release on Hugging Face includes model weights, training code, data recipes, hyperparameters, full training logs, intermediate checkpoints at steps 110k, 115k, and 162k, and evaluation code — all under the Apache-2.0 license.
Q: Can companies use Aether-7B-5Attn commercially?
A: Yes. The Apache-2.0 license permits commercial use, modification, and redistribution, making Aether-7B-5Attn freely usable by both researchers and commercial developers without additional licensing requirements.
Source: ReasonCore (영문) (2026-07-20) — original article