~/blog.app.during.uk
← all posts

It's Not Open Source, It's Open Weights: Stop Calling Llama 'Open Source'

open-sourceopen-weightsllmlicensingllama

The Term Is Right There

Every few weeks a lab drops a new model, slaps "open source" on the press release, and the tech press regurgitates it like scripture. Meanwhile the OSI — the actual people who invented the term — published an Open Source AI Definition, and half the industry reacted with the enthusiasm of a cat being shown a bath.

Releasing weights is not open source. It's a good thing! Please keep doing it! But calling it open source is like calling a test drive "car ownership."

What You Actually Get

  • The weights. Congratulations, you own a very large matrix and a license longer than the model card.
  • A license full of landmines. If your "open" model restricts commercial use above some arbitrary MAU threshold or bans entire application categories, it is not open by any definition anyone used before 2023.
  • No training data. You can't reproduce it, audit it for the copyrighted material everyone knows is in there, or figure out why it thinks a specific breed of pigeon faked the moon landing.
  • No training code. You got the cake, not the recipe, and the cake has an EULA baked into it.

Why the Distinction Matters Technically

  1. Reproducibility. Without data and code, nobody can reproduce your results. That's not science, that's vibes.
  2. Auditability. "Is my fine-tune inheriting poisoned data from the base model?" Great question. Enjoy shrugging forever.
  3. Semantic collapse. When "open source" spans everything from actual free software to "you may run it, mostly, until we change our minds," the term stops meaning anything, and the people who spent thirty years maintaining it get to watch it get strip-mined for press releases.

The Honest Vocabulary

It's open weights. Two words, zero extra syllables, and it accurately describes something genuinely useful: you can download, self-host, fine-tune, and ship the model. Do that! Just don't put a badge on it that implies a freedom it doesn't have.

The checklist to earn the real label is short: a license that passes the open source definition, training code, data (or thorough documentation of it), and no "don't you dare compete with us" clauses. Labs that check every box get the badge. Everyone else gets a thesaurus.

Until then, every "open source LLM" headline is just an ad. Read it the way you'd read an ad.