Open Weights Aren’t Open Source — and a Korean Startup Just Proved the Difference

2026-07-19 · Source: LLM on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, medium

Summary

Aether-7B-5Attn, a ~6.59B-parameter foundation model developed by Korean startup VIDRAFT, redefines "open source" in the LLM space by releasing its entire manufacturing process under an Apache-2.0 license. Unlike typical "open weights" models, Aether provides full architecture source, a byte-reproducible training data recipe, training code, hyperparameters, logs, evaluation code, and intermediate checkpoints. This comprehensive release, achieved by a single company rather than a national consortium, establishes a new standard for "sovereign AI" by enabling full verification and rebuildability. The model also features an experimental heterogeneous attention architecture, mixing five distinct attention mechanisms across 49 layers, with a 7x7 Latin square design used as an experimental control to eliminate depth bias. Trained on 144.2B tokens, its data mix notably includes 37.8% math and equal 21.6% Korean and English content, all from public datasets. Users should note an inference caveat requiring batch_size = 1.

Key takeaway

For AI Scientists and ML Engineers evaluating foundation models, Aether-7B-5Attn sets a new standard for transparency and reproducibility. If your organization prioritizes verifiable AI or seeks to build truly sovereign capabilities, you should examine VIDRAFT's comprehensive release. This model demonstrates that full control over your AI stack, including data and training processes, is achievable, moving beyond mere "open weights" to true open source. Consider adopting similar transparency for your internal or public model releases.

Key insights

True "open source AI" demands full reproducibility of training data, code, and logs, not merely published model weights.

Principles

• "Open source" implies full reproducibility, not just compiled binaries.
• AI sovereignty stems from verifiable and rebuildable models.
• Latin squares can serve as experimental controls for depth bias.

Method

A "fully open" model releases architecture, byte-reproducible data recipes, training code, hyperparameters, logs, evaluation code, and intermediate checkpoints.

In practice

• Rebuild the training corpus byte-for-byte from published recipes.
• Analyze heterogeneous attention effects using the Latin square design.
• Ensure batch_size = 1 for Aether-7B-5Attn inference to prevent output corruption.

Topics

• Open-Source AI
• Sovereign AI
• Heterogeneous Attention
• Latin Square Design
• Aether-7B-5Attn
• Model Reproducibility

Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• [AINews] Much ado about Open Weights — Moonshot AI's Kimi K3 sets a new open-weights performance bar, intensifying policy debates on model access and security.
• The Open-Source Empire’s Faith Crisis — After the Llama 4 Cheating Scandal, What’s Left of Meta’s… — Misrepresenting "open source" and manipulating benchmarks erodes trust and damages an ecosystem.
• DeepSeek STRIKES BACK - Open-Source Model Challenges OpenAI And Anthropic — DeepSeek V4 sets a 1 million token context as standard, democratizing advanced AI capabilities through open-weight MoE models.
• Welcome to Open Source AI: Run Your Own Models Locally — Running open-weight models locally offers privacy, cost savings, and performance comparable to cloud APIs through quantization and optimized inference engines.
• [AINews] "Laguna S 2.1 Released: Cheaper than Deepseek v4 Flash, Better than V4 Pro" — Open-source models are critical for defensive AI and cost-efficiency, while agent platforms mature with advanced orchestration and evaluation tools.
• TAI #199: Gemma 4 Brings a Credible US Open-Weight Contender Back to the Table — Gemma 4 offers a US-origin, Apache 2.0 licensed open-weight model family for self-hosting and regulated enterprise use cases.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by LLM on Medium.