Korean AI startup Bidraft (CEO Kim Min-sik) announced that it has released its self-developed open-source foundation model “Aether-7B-5Attn” on the global AI platform Hugging Face under the Apache-2.0 license.

This model aims to be an open-source model that includes not only the weights but also key development-process data such as the composition of training data, training code, hyperparameters, logs, and evaluation code. It is designed to enable transparent understanding of the model’s development pipeline, allowing external researchers and companies to verify its technical structure and reproduce research based on it.

Aether-7B-5Attn is a Mixture-of-Experts (MoE) architecture with a total of 6.59 billion parameters, and it is designed so that approximately 2.98 billion parameters per token are activated during inference to enhance computational efficiency. It combines five attention mechanisms—Full, Differential, Sliding Window, NSA-series sparse, and Hybrid—and arranges 49 layers in a 7×7 Latin square structure so that the various attention types are evenly applied across depths. Training was carried out on 16 NVIDIA B200 GPUs over approximately 144.2 billion tokens. The dataset is composed of 37.8% mathematics, 21.6% Korean, and 21.6% English, reflecting its training for both Korean- and English-language environments.

The disclosure of this development process may also serve as a reference case for securing technological independence and security in the field of artificial intelligence. Because the training process is public, the model can be used by research institutes and companies as basic research material to reduce external dependence and advance AI systems tailored to their own data and security requirements. As a single startup, Bidraft aims to contribute to expanding the diversity of the domestic open-source AI ecosystem by introducing a model with enhanced openness in its development process.

Bidraft is a deep-tech company that provides R&D solutions in the fields of physics, chemistry, and life sciences, based on research in quantum computing technology and AI model development. It has presented proprietary system architectures such as the Darwin model family and has achieved results in major domestic and international AI performance evaluations (benchmarks) and drug discovery-related tests using these models. It has also recorded a cumulative 1 million downloads on the open-source community platform Hugging Face, and in addition to model development, it is conducting parallel research on AI infrastructure technologies such as the inference acceleration engine VKAE and the lightweight execution engine VKUE.

Kim Min-sik, CEO of Bidraft, stated, “The foundation of the open-source ecosystem lies not only in utilizing publicly available model weights but also in creating an environment where data, code, and the entire training process can be transparently verified,” and added, “We hope that this release will serve as an opportunity to broaden constructive discussions on technological transparency and independent model development capabilities within the Korean AI research environment.”

The base and instruct models, as well as a live demo, are currently available on Hugging Face.

Popular News