VIDRAFT Releases Fully Open Foundation Model Aether-7B-5Attn: Weights, Code, Data, and Training Logs — All Public

Short summary

Korean AI startup VIDRAFT has released Aether-7B-5Attn, a ~6.59B-parameter MoE foundation model under Apache-2.0 with full reproducibility artifacts including training code, data recipes, logs, and intermediate checkpoints. The model is trained on 144.2B tokens with heavy math (37.8%) and balanced Korean/English (21.6% each) weighting. This positions it alongside OLMo and LLM-jp in the small family of genuinely open foundation models, though no benchmark scores are reported yet.

• •VIDRAFT released Aether-7B-5Attn: ~6.59B-param MoE model with 5 attention mechanisms, Apache-2.0 licensed
• •Full open release includes weights, training code, data recipes, logs, intermediate checkpoints, and eval code
• •Trained on 144.2B tokens with 37.8% math, 21.6% Korean, 21.6% English; no benchmark scores reported yet

Generated with AI, which can make mistakes.