Is AETHER free to use commercially?

Yes. AETHER's weights, data, code, logs, and checkpoints are all released under Apache-2.0, which permits commercial use, modification, and redistribution.

Where can I download AETHER?

AETHER is available on Hugging Face under the FINAL-Bench organization, including base and instruct variants with three checkpoints of about 13 GB each, and related information is on VIDRAFT's official site (vidraft.net).

Does AETHER outperform top frontier models?

No benchmark scores were published at announcement. AETHER's strength is reproducible full openness and its 7x7 Latin-square heterogeneous-attention method, not a performance ranking, and anyone can measure its performance using the included evaluation code.

VIDRAFT / Insights / Open foundation

Open foundation

What is AETHER, VIDRAFT's fully open foundation model?

A from-scratch foundation model that opens not just weights but training data, code, logs, and checkpoints — all under Apache-2.0

Published 2026-07-31About 2min readby VIDRAFT

Quick answer

AETHER is a from-scratch foundation model from Korean AI company VIDRAFT, released under a full-openness standard: the weights, training data, training code, training logs, and intermediate checkpoints are all public under Apache-2.0. This goes a step beyond the common 'open weights only' release.

What does it mean that AETHER is 'fully open' rather than just open-weight?

AETHER publishes not only its final weights but also the training data, training code, training logs, and intermediate checkpoints, all under the Apache-2.0 license.

Most 'open' models ship only finished weights, so outsiders cannot verify or reproduce how the model was trained. AETHER follows the full-openness standard set by Allen AI's OLMo, providing all five artifacts (weights, data, code, logs, checkpoints) so researchers and companies can audit the entire training process and reproduce or continue it end to end.

What is the AETHER-7B-5Attn architecture?

AETHER-7B-5Attn is a Mixture-of-Experts model with about 6.59 billion parameters that arranges five heterogeneous attention types across 49 layers using a 7x7 Latin-square layout.

The 7x7 Latin-square layout is designed so that five distinct attention types are distributed evenly across the layers, avoiding over-reliance on any single mechanism and letting the combined effect of heterogeneous attention be studied. AETHER's core value lies not in benchmark supremacy but in this reproducible attention-arrangement method and its fully open training pipeline.

What data and hardware was AETHER trained on?

AETHER was trained on about 144.2 billion tokens using 16 NVIDIA B200 GPUs, with a data mix of math 37.8%, Korean 21.6%, English 21.6%, and other 19%.

The release includes both base and instruct variants, each accompanied by three checkpoints of roughly 13 GB each. A data mix weighted most heavily toward math, with Korean and English held at equal shares, signals that the model targets reasoning ability alongside bilingual use.

What are AETHER's current limitations?

The announcement does not yet publish benchmark scores (evaluation code is included so third parties can run standard benchmarks themselves), and this build has no KV-cache, which affects latency and memory.

AETHER is a research release valued for reproducible openness and its heterogeneous-attention method, not a claim of frontier-scale performance. Rather than self-reporting benchmark numbers, VIDRAFT ships the evaluation code so anyone can measure results, and the absence of a KV-cache is an engineering detail that future builds can address.

Frequently asked questions

Is AETHER free to use commercially?: Yes. AETHER's weights, data, code, logs, and checkpoints are all released under Apache-2.0, which permits commercial use, modification, and redistribution.
Where can I download AETHER?: AETHER is available on Hugging Face under the FINAL-Bench organization, including base and instruct variants with three checkpoints of about 13 GB each, and related information is on VIDRAFT's official site (vidraft.net).
Does AETHER outperform top frontier models?: No benchmark scores were published at announcement. AETHER's strength is reproducible full openness and its 7x7 Latin-square heterogeneous-attention method, not a performance ranking, and anyone can measure its performance using the included evaluation code.

Sources

On-device AI

Can you run a 35B AI model with no GPU?

AI for Science

Which AI model is best at predicting a drug's human intestinal absorption (HIA)?

Model reasoning

What is the best Korean LLM on the GPQA Diamond science benchmark?

↖ Home — vidraft.net

This article is based on VIDRAFT public, measured data and external sources. Performance figures are measurements under the stated conditions and may vary by environment.