Open source vs open weights: three model releases, three different scopes

#machinelearning #llm #opensource #ai

Open source vs open weights: three model releases, three different scopes

"Open source model" has quietly stopped meaning one thing. Some releases hand you a .safetensors file and a paragraph of prose. Others hand you the data recipe, the training code, the logs and the checkpoints. Both get called open, and the difference between them is the difference between using a model and reproducing one.

Korean outlet ETNews covered a set of releases from VIDRAFT this week that happens to illustrate the whole spectrum — three models, deliberately published under three different scopes. This post is the map. Deep dives are linked at the bottom.

The three

	Aether-7B-5Attn	AETHER-7B-7Attn-base	Aether-6B-11Attn-base
Scope	fully open source	open weights	open weights
Weights	✅ Apache-2.0	✅ Apache-2.0	✅ Apache-2.0
Architecture source	✅	✗ proprietary	✗
Training code	✅	✗	✗
Data recipe	✅	✗	✗
Logs	✅ all 162,000 steps	✗	✗
Intermediate checkpoints	✅	✗	✗
Parameters	6.59B / ~2.98B active	6.59B / ~3B active	5.79B
Stage	final base	base	mid-training (step 174,000)

Three things worth stating plainly, because release pages tend not to:

1. Only one of these is open source. Aether-7B-5Attn ships the recipe, the code, the hyperparameters, the logs and the checkpoints. The other two ship weights. All three are Apache-2.0 on the weights — which is a licence statement, not a reproducibility statement. Conflating those two is the single most common way "open" gets oversold, and it is worth checking on every release you evaluate, including these.

2. AETHER-7B-7Attn-base and Aether-7B-5Attn are different checkpoints. I checked this at the byte level rather than taking a card's word for it: same file size (13,179,635,998 bytes), same safetensors header, identical tensor names, shapes and offsets — but different SHA-256, and every sampled tensor block differs. Same architecture family, different training state. If you are benchmarking, they are not interchangeable.

Aether-7B-5Attn        sha256 8e77ad511a524347dd1c…
AETHER-7B-7Attn-base   sha256 9250f1f8a51ed6f3593a…

3. The 11Attn release is still training. Step 174,000 is a mid-flight snapshot, not a finished model. It is published as a research artifact and the card says so.

What the shared design is actually for

The one architectural idea common to all three is public because the open-source member documents it: layers do not all have to run the same attention mechanism.

That uniformity — every layer running the same mixer — is a convention inherited from the reference implementation, not a proven optimum. The obvious experiment is to put several mechanisms in one network. The non-obvious problem is that the naive version proves nothing.

Transformer layers are not interchangeable; early layers do different work than late ones. Stack mechanisms in some order and any measured difference is confounded — you cannot separate "this mechanism is good" from "this depth is where the work happened." That is depth bias, and it is enough to make a result uninterpretable.

The fix used here is a Latin square: a grid where each symbol appears exactly once per row and per column, so every mechanism lands at every depth position exactly once. The architecture is an experimental control before it is anything else.

The layout as published with the open-source member of the family. You can only see a diagram like this — let alone check it against the code — when the architecture source ships with the weights.

I verified how well that control actually holds — and it is more interesting than the marketing version. Details in the 5Attn audit, where the code to check it is published.

The claim nobody here is making

That heterogeneous mixing beats a uniform stack.

No such result exists yet, for 5, 7 or 11 mechanisms. Establishing it needs a controlled ablation — same data, same token budget, same compute, heterogeneous vs. homogeneous — and until that is run the honest status is unproven.

That sentence is uncomfortable to publish, and it is the reason the open-source member of the family exists at all. A hypothesis nobody else can test is not a scientific claim; it is marketing. Aether-7B-5Attn ships everything needed to run the experiment — including everything needed to produce a result that contradicts the people who published it.

The other two releases are weights: useful to run and benchmark, not instruments for that experiment. Knowing which is which before you build on one is the entire practical point of this post.

If you load any of them

batch_size = 1. Sparse branches in this family do not consume the padding mask. Left-pad a batch and queries can attend into pad tokens — no exception raised, just quietly wrong output.
Base checkpoints repeat under greedy decoding. Expected for un-tuned bases; use sampling with a repetition penalty. An instruct variant of the open-source model is published separately.
The 11Attn artifact is slow and un-tuned — treat it as research, not a serving build.