THE VERDICT
● Open weights (Kimi K3, GLM-5.3, Qwen) for capability. These are the stronger models.
● Fully open (K2 Horizon) where you need to prove what went in — regulated work, research, anything with an audit trail.
● The honest position: most projects do not need provenance, and the ones that do cannot substitute for it.
What each actually gives you
| Published | Open weights | Fully open |
|---|
| Weights | Yes | Yes |
| Training code | Rarely | Yes |
| Training data | No | Yes |
| Checkpoints and logs | No | Yes |
| Frontier capability | Closer | Not claimed |
FOUR QUESTIONS YOU CANNOT ANSWER WITHOUT THE DATA
Was copyrighted material used, and how was it obtained? Is my domain represented, or is the model guessing? What biases came from the corpus rather than the tuning? Could I reproduce this if the lab disappeared?
Open weights answer none of these. For a side project that is fine. For a regulated deployment it may not be.
Why the first question is not theoretical
A federal court has already held that training on copyrighted material can be lawful while acquiring it through piracy is not, and ordered $1.5 billion. Sony and Warner sued Anthropic on 28 August on exactly that question.
If you fine-tune on a base model whose corpus is undisclosed, you cannot establish what is in your derivative. That is not a reason to avoid open-weight models — it is a reason to know which question you cannot answer.
Which one
| If you are... | Pick |
|---|
| Building a product on capability | Open weights. Kimi K3 has the clearest terms of the large ones |
| Working in a regulated sector | Fully open. Provenance is the requirement, not the bonus |
| Doing research on training dynamics | Fully open. Checkpoints and logs are the whole point |
| Fine-tuning for a client | Check what your contract says about warranting the training corpus |
FAQ
Are fully open models worse?
On capability, generally yes at the moment — the largest open-weight models are stronger. Full openness is a different axis and does not imply performance.
Does open weights mean open source?
No, and that is the Open Source Initiative's objection. Weights alone cannot be inspected or reproduced.
Does any of this affect me if I just use an API?
No. This matters if you self-host, fine-tune, or need to establish provenance for compliance.