Not "open weights." Genuinely open — weights, training-data recipe, code, every hyperparameter, and the complete logs, all released under Apache-2.0. A 6.59B heterogeneous-attention MoE, built from scratch by VIDRAFT.
Answers here come from Aether-7B-5Attn-it, the instruction-tuned model — not the base. The base (Aether-7B-5Attn) is the reproducible artifact this release is about; it is not instruction-tuned and would not answer questions like this.
Serving Aether-7B-5Attn-it (6.59B MoE) on a single T4, at batch_size=1. This build ships no KV cache, so generation is slow and capped — tokens stream in as they are produced.
A 6.59-billion-parameter Mixture-of-Experts language model (~2.98B active per token) whose 49 layers place five distinct attention mechanisms on a 7×7 Latin square, so each mechanism appears exactly once at every depth. Trained from scratch on 144.2B tokens (≈37.8% math, 21.6% Korean, 21.6% English, 13.5% code) and released fully open under Apache-2.0.
| Total / active params | 6.59B / ~2.98B | Layers | 49 (7×7 Latin square) |
|---|---|---|---|
| Experts | 25, top-7 + 1 shared | Context | 4096 |
| Tokens | 144.2B | License | Apache-2.0 |
One architecture line, released at different levels of openness.
Yes — 100% open. Weights + training-data recipe + code + all hyperparameters + complete logs + full architecture source, under Apache-2.0. This is the "fully open" standard, not "open weights only."
VIDRAFT (주식회사 비드래프트), a Korean AI startup — the only single startup among the world's sovereign fully-open foundation models (alongside OLMo, Soofi, Apertus, LLM-jp, Viking).
It is the only sovereign fully-open model with a self-designed attention layout: five mechanisms (full, differential, sliding, NSA, hybrid) on a 7×7 Latin square — a control that removes depth bias.
Yes. Every source is a public repository; the released recipe reproduces the training binaries byte-for-byte.