OPEN
ENOUGH
what you get when you get the weights
You can download the weights. You cannot download what made them. The word open covers both, which is why the arguments never end.
IT professional exploring how AI systems work under the hood.
01 WEIGHTS ARE NOT SOURCE
the parameters, not the recipe
An open-weights release hands you a file of numbers. Parameters, a config, a tokenizer. That is enough to run the model on your own hardware and enough to fine-tune it.
It does not hand you the training corpus, the code that did the training, or the checkpoints along the way. You can use the result. You cannot rebuild it, and you cannot audit what went into it.
Software has a word for the left column. It calls it a binary. Nobody calls a binary open source.
02 FOUR RUNGS, NOT TWO CAMPS
the word is doing a lot of work
There is no open and closed. There is a ladder, and most arguments are two people standing on different rungs using the same word.
Closed means no weights at all — your text goes to someone else's machine and an answer comes back. Restricted means you can download and run it, inside limits the licence sets. Permissive means Apache-2.0 or similar, with no field-of-use cap. Fully open means the weights, the training code, and enough detail about the data to rebuild the thing.
Only the last rung lets a stranger check your work.
03 READ THE LICENCE, NOT THE HEADLINE
open with conditions is a different product
The Open Source Initiative published a definition on October 28, 2024. To qualify, a release needs the training and inference code, the parameters, and enough information about the data that a skilled person could substantially recreate the system. It stops short of requiring the dataset itself — which the Free Software Foundation and the Software Freedom Conservancy both said was too weak.
By that definition Llama does not qualify. Section 2 of its licence says that once you pass 700 million monthly active users, you need Meta's written permission to carry on. Meta's position is that no single definition fits models like these.
Both things are true at once. It is genuinely downloadable. It is not open source.
04 THE QUESTION THAT SURVIVES
not which is better. what happens when it changes.
Benchmarks move. Whichever side leads this quarter will not lead forever, and choosing on capability alone means choosing again every few months.
The durable question is control. If the vendor retires the version you built on, what breaks. If you have to show two years from now what produced a given output, can you. If your data cannot leave the building, the question answered itself before you asked it.
Closed buys capability and removes the operations. Open weights buy a version that cannot change under you, and hand the operations back.