Korean startup VIDRAFT released Aether-7B-5Attn on July 19, 2026, a 6.59-billion-parameter language model published with full training data recipes, code, logs, and checkpoints under an Apache-2.0 license. The model experiments with five distinct attention mechanisms arranged in a 7×7 Latin square across 49 layers, testing whether different computational approaches benefit different depths of the network. It was trained on 144.2 billion tokens using 16 NVIDIA B200 GPUs, and the team describes it as the first fully open sovereign foundation model built by a single AI startup rather than a national institution.
