The terms open weights and open source get used as if they mean the same thing. They do not, and the difference can affect what an Australian business is legally allowed to do with a model. A licence you never read can sit underneath a product you have already shipped.
With the wave of open model releases this year, plenty of Australian teams are weighing self-hosted options against managed models like Claude. Before any of that, it pays to be precise about what the open label on a model actually grants you. Here is the plain version, without the licence-lawyer jargon.
What open weights actually gets you
When a lab publishes open weights, you receive the trained model parameters: the big file of numbers you can run on your own hardware. That is genuinely useful, but it is the finished cake, not the recipe.
You get the trained parameters and usually some inference code
You almost never get the training data
You rarely get the full training code or methodology
Usage terms come from a custom licence, not your ownership of the file
Running the model locally feels like ownership. Legally, you are a licensee, and the licence is the document that decides what you can do.
What open source means, and how rarely it applies
Open source is a stricter idea borrowed from software. A true open source release implies the freedom to use, study, modify and redistribute, under a recognised licence such as Apache 2.0 or MIT. For AI models, the honest picture is messier.
Very few models ship with training data, code and weights all open
Many popular open models use custom licences with commercial restrictions
Some licences cap usage by company size or user count
Terms can change between versions of the same model family
The label on the announcement is a poor guide to your actual rights. Two models described the same way in a headline can carry very different obligations.
Why the difference matters commercially
For a business, this is not an ideological debate. The licence shapes your rights, your risks, and what you can promise customers.
Some open-weight licences restrict or condition commercial use
Without training data, you cannot fully audit bias or provenance
Indemnity is usually absent: if the model causes harm, that risk sits with you
A licence change upstream can strand a product you have already built
If your product touches personal information, the Privacy Act obligations stay yours regardless of what the model licence says. Self-hosting an open model does not transfer compliance risk to the lab that released it.
Reading the licence, not the label
For an Australian SMB, the sensible move is a short, boring process rather than trust in the word open. A focused legal review of a model licence typically costs $3,000 to $5,000. That is cheap insurance against discovering that a core product rests on terms you cannot use, where a rebuild can run past $50,000 once engineering time is counted.
Check whether commercial use is genuinely permitted for your case
Note exactly what you receive: weights, code, data, or weights alone
Keep a register of which model versions you use and their terms
Re-check the licence whenever you upgrade model versions
Default to clearly licensed or managed options for revenue-critical products
Common mistakes to avoid
Most licence trouble is avoidable. The same few errors show up again and again when teams move fast on a model choice.
Treating open weights as if it meant public domain
Assuming the licence on version 2 still applies to version 3
Building a customer-facing product before anyone reads the terms
Forgetting that fine-tuned derivatives usually inherit the base licence
Letting each team adopt models with no shared register of what is in use
What this means for Australian businesses
A Sydney or Melbourne firm choosing between a self-hosted open model and a managed model like Claude is really choosing between two risk profiles. With open weights, you own the infrastructure and the legal homework. With a managed model, the provider carries the licence clarity and you pay per use. Neither is wrong, but the choice should be made with the licence on the table, not the press release.
We read the licence terms before any model goes near production
We keep a model register for every client deployment
We default to Claude where licence clarity matters to the business
Key takeaways
Open weights means you get the parameters, not the recipe
Open source is a stricter standard that few AI models truly meet
The licence, not the label, decides what your business can build
A $3,000 to $5,000 review is cheap against a stranded product
Talk to a Claude specialist
Automata AI is a Sydney based consultancy that helps Australian businesses put AI to work on clear legal footing. If you are weighing open models against Claude for a real workload, book a short brainstorm and we will map the trade-offs for your situation.
