Aether-7B-5Attn: VIDRAFT Open-Sources Full Training Stack — Weights, Code, Hyperparameters, and Logs
TL;DR: Korean AI startup VIDRAFT (Bidraft) has released
Aether-7B-5Attnon Hugging Face under the Apache 2.0 license — a 7B-parameter foundation model that goes beyond typical open-weight releases by also publishing the complete training source code, data recipe, hyperparameters, training logs, intermediate checkpoints, and evaluation code. If you care about reproducible AI research or want a genuinely fully-open base model, this one is worth your attention.
What it is
Aether-7B-5Attn is a 7-billion-parameter open-source foundation model developed by VIDRAFT (also transliterated as Bidraft), a Seoul-based deep-tech AI company. It is released on Hugging Face under the Apache 2.0 license, which means commercial use, modification, and redistribution are all permitted.
What sets this release apart from the crowded field of "open-weight" models is the scope of what's actually being opened:
- Model weights — the trained parameters themselves
- Training data recipe — the description of what data was used and how it was mixed/processed
- Full training source code — the actual code used to train the model
- Hyperparameters — the configuration used during training runs
- Training logs — step-by-step records of the training process
- Intermediate checkpoints — snapshots from various points during training, not just the final model
- Evaluation code — the scripts used to benchmark the model
This level of disclosure is what the research community sometimes calls a "fully open" model, as opposed to models that release weights only while keeping training pipelines proprietary. The naming suffix 5Attn appears to reference an architectural detail of the attention mechanism, though specifics beyond the model name have not been disclosed publicly.
How it works
Aether-7B-5Attn is part of VIDRAFT's broader AETHER architecture family, which the company describes as the "design the future" layer within its four-tier integrated AGI system. The other tiers include the Darwin model family (production-facing), PROMETHEUS (a world model), and HEPHAESTUS (embedded AI).
At a conceptual level, the 5Attn designation in the model name suggests a specific configuration of the attention layers — a common axis of differentiation among modern transformer-based LLMs. Beyond that, the complete training code and hyperparameter files published in the repository are the canonical reference for how this model was built. Rather than relying on a paper or a blog post, engineers can inspect the actual training pipeline directly.
The emphasis on releasing intermediate checkpoints is particularly useful for the research community: it enables studying training dynamics, loss curves, capability emergence, and the effect of data curriculum over the course of training — not just the final artifact.
Benchmarks & results
The source article references several public benchmark results associated with VIDRAFT's Darwin model family (the production tier, of which AETHER is the underlying architecture):
- K-AI Leaderboard: Darwin models ranked #1 overall on this Korean AI evaluation leaderboard
- GPQA Diamond: Score of 90.9% — GPQA Diamond is a graduate-level science Q&A benchmark widely used to assess expert-level reasoning
- Polaris Global New Drug Development Leaderboard: 14 category wins — Polaris is a domain-specific benchmark focused on drug discovery and life sciences
- Metacognition Leaderboard: Ranked #1
- Hugging Face downloads: The Darwin/AETHER model family has surpassed 1 million cumulative downloads on Hugging Face
Note that these figures are reported for the Darwin/AETHER family broadly. Benchmark results specific to Aether-7B-5Attn individually should be verified against the model card on Hugging Face.
How to try it
The model is publicly available on Hugging Face. You can download it using the standard Hugging Face CLI:
pip install huggingface_hub
huggingface-cli download bidraft/Aether-7B-5Attn
Note: The exact repository path should be confirmed on the VIDRAFT Hugging Face profile — use the repository slug as listed there.
Because the full training code and evaluation scripts are also published, you can clone the associated repository to reproduce training runs or run evaluations:
# Check the Hugging Face model card for the linked GitHub/code repository URL
git clone <repository-url-from-model-card>
The source article does not mention a public OpenAI-compatible inference API endpoint for this specific model at this time. For inference, the standard approach would be loading the weights locally via transformers or a compatible inference framework.
FAQ
Q: Is this actually "fully open" or just open-weight like most models?
A: Based on the disclosure, this goes significantly further than open-weight-only releases. VIDRAFT is publishing training source code, data recipes, hyperparameters, logs, intermediate checkpoints, and evaluation code alongside the weights — which qualifies as a genuinely fully-open release by most community definitions.
Q: Can I use this for commercial projects?
A: Yes. The Apache 2.0 license permits commercial use, modification, and redistribution. Always review the full license text and any supplementary terms on the Hugging Face model card before deployment.
Q: What does the "5Attn" in the model name mean?
A: The suffix appears to reference a specific attention architecture configuration. The training source code and model configuration files published in the repository are the authoritative reference for this architectural detail.
Q: How does Aether-7B relate to VIDRAFT's Darwin models?
A: AETHER is VIDRAFT's architecture layer — described by the company as "designing the future" — while Darwin is the production-facing model family built on top of it. Aether-7B-5Attn is a foundational release from the AETHER tier.
Originally reported by 金融界 (중국) (2026-07-20) — source article.
