Korean Startup Releases Aether-7B, Publishing Even Its Training Data

A Korean AI startup has released Aether-7B-5Attn on Hugging Face under the Apache-2.0 license, disclosing not just the weights but also the training data recipe, architecture source, and training logs. This goes beyond a simple open-weight model, representing a fully open-source model to the extent that a third party could reproduce the same training binary.

The key technical features of this model are as follows:

• Variable attention mechanism: Departing from the convention of using the same attention across all layers, it deploys 5 different attention types (Full, Sliding, Differential, NSA, Hybrid) depending on each layer's role.
• Efficient compute design: The cost profile of each mechanism was measured and published; for example, Sliding Attention shows roughly 1.8x faster speed compared to Full Attention at a 32K context.
• Ensuring reproducibility: Data sources, tokenizer, mixture weights, and the full 162,000-step training logs are included, maximizing the model's transparency.

This release is particularly significant in that it was achieved not by a national research institute with large-scale infrastructure, but by a single startup using 16 B200s.