A Seoul-based Korean AI startup makes its entire training pipeline public — weights, data, code, and all.
TL;DR: VIDRAFT, a Korean AI startup headquartered in Seoul, has released Aether-7B-5Attn as a fully open-source foundation model 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, training logs, intermediate checkpoints, and evaluation code. The model features a mixture-of-experts architecture with 6.59 billion total parameters and five distinct heterogeneous attention mechanisms, and is designed as a sovereignty-first, bilingual Korean-English model rather than an English-centric one.
VIDRAFT, the Seoul-based Korean AI startup pursuing Pre-AGI technology, published its Aether-7B-5Attn foundation model on Hugging Face on July 20, 2026 — making the release one of the most transparently documented open-source model drops to come out of the Korean AI ecosystem this year. Chinese IT publication CNMO 手机中国 spotlighted the launch, describing Aether-7B-5Attn as an "autonomous foundation model" that places Korean and English on equal footing rather than defaulting to the English-first paradigm common across most large language models today.
The centerpiece of the release is Aether-7B-5Attn, a mixture-of-experts (MoE) model with 6.59 billion total parameters. To keep inference efficient, only approximately 2.98 billion parameters are activated per token — a design choice that reduces compute cost during deployment without sacrificing the representational capacity of the full parameter set.
What sets Aether-7B-5Attn apart architecturally is its integration of five distinct attention mechanisms within a single model: Full Attention, Differential Attention, Sliding Window Attention, NSA-series sparse attention, and a Hybrid mode. The 49-layer network arranges these mechanisms in a 7×7 Latin square pattern, deliberately distributing each attention type across different network depths to prevent any single mechanism from dominating a particular depth range. VIDRAFT describes this as avoiding the clustering of specific attention strategies in specific layers.
On the data side, the model was trained on approximately 144.2 billion tokens. The training corpus was intentionally balanced: mathematics accounts for 37.8% of the data, with Korean and English each comprising 21.6% — a deliberate choice that positions Aether-7B-5Attn as a bilingual, sovereignty-conscious model rather than one optimized primarily for English-speaking use cases.
The full release package available on Hugging Face includes the base model, an instruction-tuned variant, and a live demo. All training code, the complete data recipe, training logs, intermediate checkpoints, and evaluation code are also publicly accessible under the Apache 2.0 license, which permits both commercial and non-commercial use.
Beyond Aether-7B-5Attn itself, CNMO's coverage notes that VIDRAFT operates across a broader four-tier integrated AGI system composed of the Darwin model family, the AETHER architecture, the PROMETHEUS world model, and HEPHAESTUS embedded AI. The company has also developed proprietary AI infrastructure, including inference acceleration and lightweight runtime engines. The Darwin model family has surpassed one million cumulative downloads on Hugging Face, and the series has recorded benchmark achievements including a top ranking on the K-AI leaderboard, a 90.9% score on GPQA Diamond, fourteen first-place finishes on the Polaris global drug discovery leaderboard, and a top position on the metacognition leaderboard.
Full transparency releases — where a team publishes not just weights but the complete recipe for reproducing a model from scratch — remain relatively rare in the foundation model space. Most open-weight releases still withhold training data details or code, making genuine reproducibility difficult. VIDRAFT's decision to expose every layer of the pipeline, from raw data composition to intermediate checkpoints, puts Aether-7B-5Attn in a small category of models that researchers and developers can fully audit and rebuild.
The bilingual parity between Korean and English in the training data is also noteworthy in a broader geopolitical context. As nations increasingly seek technological self-sufficiency in AI, a foundation model that treats a non-English language as a first-class citizen — rather than a secondary fine-tuning target — represents a meaningful step toward linguistic and strategic sovereignty in AI development.
For the open-source AI community, the architectural novelty of combining five attention types within a structured Latin square layer arrangement offers a concrete research artifact to study how heterogeneous attention strategies interact at scale.
Q: What is Aether-7B-5Attn and who made it?
A: Aether-7B-5Attn is a fully open-source foundation model developed by VIDRAFT, a Korean AI startup based in Seoul. It was released on Hugging Face under the Apache 2.0 license on July 20, 2026.
Q: What makes Aether-7B-5Attn different from other open-source models?
A: The model combines five distinct attention mechanisms arranged in a 7×7 Latin square across its 49 layers, uses a bilingual Korean-English training corpus with equal weight given to both languages, and releases the complete training pipeline — including data recipe, code, logs, and checkpoints — for full public reproducibility.
Q: Is Aether-7B-5Attn free to use commercially?
A: Yes. The model is released under the Apache 2.0 license, which permits both commercial and non-commercial use without restriction.
Source: CNMO 手机中国 (중국) (2026-07-20) — original article