aiAI타임스 (AI Times)· 2026. 7. 20. 오전 3:01:078.0

BeDraft Unveils Comprehensive Open-Source Model Aether-7B-5Attn

AI specialist BeDraft (CEO Kim Min-sik) announced on the 20th that its open-source foundation model 'Aether-7B-5Attn' has been released on the global AI platform HuggingFace. Aether-7B-5Attn, released under the Apache-2.0 license, includes not only model weights but also training data recipes, full training code, hyperparameters, learning logs, intermediate checkpoints, and evaluation code, embodying a fully open model. BeDraft, which conducts R&D across physics, chemistry, life sciences, and pharmaceuticals using AI and quantum computing technologies, presents a four-layer integrated AGI system featuring the Darwin model family, AETHER architecture, PROMETHEUS world model, and HEPHAESTUS embedded AI. The company's direction, stated on its website, is 'Darwin creates results, AETHER designs the future.' BeDraft has previously showcased achievements such as ranking first in the K-AI leaderboard for Darwin series models in 2026, achieving 90.9% on the GPQA diamond, leading the Polaris global new drug development leaderboard with 14 wins, and ranking first on the meta-cognition leaderboard. The Darwin model family has also surpassed 1 million cumulative downloads on HuggingFace. Beyond model development, the company has also developed AI infrastructure technologies like the inference acceleration engine VKAE and the lightweight runtime engine VKUE. Aether-7B-5Attn focuses on surpassing the limitations of previous models that only shared weights, designed to allow tracking of training data composition, tokenization processes, training code, configuration values, logs, and evaluation procedures through open repositories, thereby publicly disclosing the 'manufacturing process' of the model itself. The company described this as 'an attempt aligned with the open foundation model movement, emphasizing reproducibility, similar to the OLMo project by the Allen AI Research Lab.'

💡 AI 분석: 지식 공유 및 글로벌 AI 협력에 기여하는 전략적 시도
원문 보기 (AI타임스 (AI Times)) →