Korean startup VIDRAFT open-sources Aether-7B-5Attn with full training pipeline

South Korean AI startup VIDRAFT has released Aether-7B-5Attn, a roughly 6.59-billion-parameter Mixture-of-Experts language model, on Hugging Face under the permissive Apache-2.0 license. Unlike most models labeled open-source, this release goes beyond weights to include training code, data recipes, hyperparameters, training logs, intermediate checkpoints, and evaluation code, making the entire pipeline independently reproducible. The model was trained on approximately 144.2 billion tokens, with training data weighted heavily toward mathematics (37.8%), Korean (21.6%), and English (21.6%). VIDRAFT frames the release as an expression of 'Sovereign AI,' arguing that genuine AI independence requires the ability to verify and audit a model's full development process, not just download pre-trained weights. The company places Aether-7B-5Attn in the same category as fully open models from Allen AI and LLM-jp, though no specific benchmark scores were included in the announcement.