VIDRAFT Releases Aether-7B-5Attn: A Fully Open Foundation Model with Training Recipes, Logs, and Checkpoints

TL;DR: VIDRAFT, a Korean Pre-AGI AI startup, has released Aether-7B-5Attn on Hugging Face under the Apache-2.0 license — not just the weights, but the full training recipe, training code, hyperparameters, logs, intermediate checkpoints, and evaluation code. Think OLMo-style radical transparency, built in Korea. If reproducibility and auditability matter to your research or production stack, this release is worth a close look.

What it is

Aether-7B-5Attn is a 7-billion-parameter open foundation model developed independently by VIDRAFT (비드래프트, CEO Kim Min-sik). It was published to Hugging Face on July 20, 2026, and is licensed under Apache-2.0, meaning it can be used for both research and commercial purposes without royalty obligations.

What distinguishes this release from a typical open-weight drop is its scope:

• Model weights — the trained parameters, ready for inference or fine-tuning
• Training data recipe — how the pretraining corpus was composed and tokenized
• Full training code — the actual scripts used to run the training job
• Hyperparameters — the configuration values used during training
• Training logs — loss curves and run artifacts so you can trace what happened during training
• Intermediate checkpoints — snapshots at multiple points in the training run, enabling checkpoint analysis or continued training
• Evaluation code — the scripts used to benchmark the model, so results are independently verifiable

VIDRAFT explicitly positions this as a move toward the fully open foundation model philosophy — the same ethos championed by the Allen Institute for AI's OLMo family — where the manufacturing process, not just the final artifact, is open.

The model sits within VIDRAFT's broader AETHER architecture tier, one layer of a four-tier integrated AGI system the company describes as: the Darwin model family, the AETHER architecture, the PROMETHEUS world model, and HEPHAESTUS embodied AI.

How it works

At a conceptual level, Aether-7B-5Attn follows the fully open training paradigm pioneered by projects like OLMo:

1. Documented data pipeline — the tokenization strategy and corpus composition are recorded and published, so researchers can understand exactly what data shaped the model's priors.
2. Reproducible training — by publishing both the training code and the exact hyperparameter configuration, any team with sufficient compute can attempt to replicate the run from scratch.
3. Checkpoint transparency — intermediate checkpoints let the research community study how model capabilities emerge and evolve across training, a practice that is uncommon in standard open-weight releases.
4. Auditable evaluation — publishing the evaluation harness alongside benchmark numbers closes the loop: you can verify the reported scores rather than taking them on faith.

The "5Attn" suffix in the model name is part of VIDRAFT's architectural naming convention for this model. The specific architectural details of what "5Attn" denotes are not elaborated in the current public announcement.

Benchmarks & results

VIDRAFT has previously published several public benchmark results for its Darwin model family (the inference/application tier above AETHER). These results have been reported as:

• K-AI Leaderboard overall #1 (2026, Darwin-family models)
• GPQA Diamond: 90.9% — a graduate-level science reasoning benchmark, positioning it among strong performers globally
• Polaris Global Drug Discovery Leaderboard: 14 consecutive #1 rankings — a specialized scientific benchmark for pharmaceutical AI
• Metacognition Leaderboard: #1
• Hugging Face cumulative downloads: 1 million+ across the Darwin model family

Note that these benchmark figures are associated with the broader Darwin/AETHER model family as reported publicly. Specific evaluation results scoped exclusively to Aether-7B-5Attn as a standalone release should be verified against the evaluation code and logs in the Hugging Face repository.

How to try it

The model is publicly available on Hugging Face. You can pull it with the standard Hugging Face CLI:

pip install huggingface_hub
huggingface-cli download vidraft/Aether-7B-5Attn

⚠️ Note: The exact repository path on Hugging Face was not specified in the source article. Search for Aether-7B-5Attn or VIDRAFT on huggingface.co to locate the canonical repository, and verify the Apache-2.0 license on the model card before use.

The repository is expected to include training code, evaluation scripts, and checkpoint links — so cloning or downloading the full repo (not just the weights) is recommended if you intend to reproduce training or benchmark results.

FAQ

Q: How is this different from other open-weight 7B models like Mistral or LLaMA?
A: Most open-weight releases publish only the final model weights — you can run inference and fine-tune, but you cannot reproduce the training run. Aether-7B-5Attn also publishes the training data recipe, code, hyperparameters, logs, and intermediate checkpoints. This puts it closer to the OLMo standard of full openness, which is rare at any scale.

Q: Is Apache-2.0 genuinely permissive for commercial use?
A: Yes. Apache-2.0 allows use, modification, and redistribution — including in commercial products — as long as you include the original license notice and attribution. There are no viral copyleft requirements. Always do your own legal review for your specific jurisdiction and use case.

Q: What is the relationship between the Darwin models and Aether-7B-5Attn?
A: VIDRAFT describes a four-tier AGI architecture. Darwin is the model family tier (deployed, inference-facing models). AETHER is the architecture tier — Aether-7B-5Attn is a model within this architectural layer, sitting below Darwin in the stack conceptually. Think of Darwin as the application-level family and AETHER as the foundational architecture that underlies it.

Q: Can I use the published training code to continue pretraining from an intermediate checkpoint?
A: Based on the stated scope of the release — which includes intermediate checkpoints and training code — this should be possible in principle. Verify the specific checkpoint formats and code documentation in the repository for practical details.

Originally reported by AI타임스 (2026-07-20) — source article.