What "Fully Open" Really Means for an LLM — and Why Aether-7B Just Landed on a Free-Models List
A small but telling thing happened this week. VIDRAFT's Aether-7B-5Attn was added to awesome-free-models, a community-curated list of AI models and tools you can use without paying. In the same update it sits next to DeepSeek V4 Pro, Mistral Large 3, Nemotron 3 Super, and Phi-4 — the newest wave of open models from far larger labs.
It's one line in a GitHub list, not a benchmark trophy. But the reason it's there is worth unpacking, because it points at a distinction the word "open" usually hides.
"Open weights" is not "open source"
Most models people call "open" are open-weight: you get a file of trained parameters you can download and run. That's genuinely useful — but it's the output, not the method. You can't see what data produced it, what code trained it, or what the training logs looked like. You can run it; you can't reproduce it, audit it, or fully trust its provenance.
The awesome-free-models entry describes Aether this way:
"100% open foundation model (weights, data, code, logs). 7B MoE (~3B active) with heterogeneous attention. Apache 2.0."
The part that matters is the parenthetical: weights, data, code, logs. Not just the trained artifact, but the whole chain that made it. That's the difference between "here's a binary" and "here's the recipe, the ingredients, and the kitchen notes."
Why a developer should care
If you build on top of models, full openness buys you things open weights alone can't:
- Reproducibility — you can retrace how the model was trained, not just guess.
- Auditability — data and logs are inspectable, which matters for compliance and for trusting what's inside.
- Sovereignty — you can rebuild or adapt it end-to-end, without depending on a vendor's black box.
- Real Apache-2.0 freedom — use it, modify it, ship it commercially, no strings.
For a foundation model — the thing everything else is built on — that transparency isn't a nice-to-have. It's the whole point of calling it "open."
What Aether-7B-5Attn actually is
Aether is a from-scratch foundation model, not a fine-tune of someone else's base. It's a 7B sparse Mixture-of-Experts design that keeps only about 3B parameters active per token, so it stays efficient to run, and it uses a heterogeneous (multi-type) attention stack rather than a single uniform attention. It ships under Apache-2.0, with the training data, code, and logs published alongside the weights.
The honest framing
This is a single entry on a community list, curated by an individual, not an official ranking — and I'd rather say that plainly than dress it up. What's meaningful is who put it there: a third party, who chose to describe Aether by the one property VIDRAFT has been building toward — being 100% open, all the way down.
That's the version of "open" worth aiming for. If you're evaluating models for something you actually have to trust, it's worth asking each one the same question: open weights, or open everything?
Aether is built by VIDRAFT — a Pre-AGI AI company working on foundation models, on-device inference, and applied AI tools.
Top comments (0)