Open Weights vs Closed Models

Models are described as open or closed, and the terms are used loosely. The distinction that matters for a user is who holds the weights and what you may do with them.

The categories

Closed, API only. The weights are the provider's. You send text, you get text back. Most of the well-known assistants.

Open weights. The model file is published and downloadable. You can run it on your own hardware, inspect it, modify it, and in many cases deploy it commercially — subject to the licence.

Open source, strictly. Weights plus training data plus training code, under a recognised open source licence. Rare. Most "open source" models are open weights with restrictions, which is a meaningful difference the marketing obscures.

Check the licence rather than the word. Some open-weight licences restrict commercial use above a size threshold, prohibit certain applications, or require attribution.

What open weights buy you

Data does not leave your control. For confidential documents, regulated data, or anything under a duty of confidentiality, this is frequently the entire argument. No third party processes the content.

No dependence on a vendor's decisions. A closed model can be deprecated, repriced, restricted or changed underneath you. A model file you hold does not change.

Reproducibility. The same model version gives the same behaviour indefinitely. For anything that must be auditable, this matters.

No per-token cost. You pay for hardware and electricity. At high volume this inverts the economics; at low volume it does not.

Offline operation.

Inspection and modification. Fine-tuning on your own data, adjusting behaviour, examining what the model does.

What they cost you

Capability gap. The strongest open models trail the strongest closed ones, though the gap has narrowed considerably and for many tasks it no longer matters.

Hardware. Useful models need substantial memory. The largest need more than a single consumer machine provides.

Operational work. Serving, updating, monitoring, scaling. This is a real engineering cost that comparisons routinely omit.

No safety layer by default. Closed products include filtering and refusal behaviour. Open weights may or may not, and the responsibility becomes yours.

Slower on modest hardware. A model that runs on a laptop runs at laptop speed.

Choosing between them

Use closed when: you want maximum capability, low operational effort, and the data is not sensitive. This covers most individual use.

Use open weights when: confidentiality is a requirement, volume is high enough that per-token pricing hurts, you need reproducibility, you need offline operation, or you need to modify the model.

Use both — the common mature arrangement. Sensitive work locally, everything else through an API.

The privacy nuance

Running a model locally means the text does not leave your machine. That is the strongest privacy position available and it is genuinely different in kind from a policy promise.

But check what else the software does. A local model served by an application that phones home, logs prompts, or syncs history is not the same as a local model. Read what the wrapper does, not just what the model is.

And a closed provider's enterprise tier is often adequate. Contractual commitments not to train on your data, regional processing, and retention controls cover many confidentiality requirements without the operational burden. The choice is not always local-or-nothing.

Hardware, briefly

Memory is the binding constraint, not processing speed. A model must fit in memory to run at reasonable speed.

Quantisation — storing weights at reduced precision — cuts memory requirements substantially with modest quality loss. This is what makes capable models run on consumer hardware at all.

Unified memory architectures are unusually well suited, because the model can use system memory rather than being limited to a graphics card.

A machine with 16 GB runs small models usefully. 32 GB opens up mid-sized ones. Beyond that you are into workstation territory.

The direction

The gap between open and closed has narrowed at every model generation, and small models have improved faster than large ones. Tasks that required a frontier model two years ago now run acceptably on a laptop.

That trend is the strongest argument for learning to work with local models even if you currently use a hosted assistant: the category of work you can do privately keeps growing.

Reading a model licence

"Open" covers licences with materially different terms, and the differences matter if the use is commercial.

Genuine open source licences — permissive or copyleft — allow use, modification and redistribution without restriction on purpose. A minority of published models.

Custom licences with use restrictions are more common. Typical clauses prohibit specified applications, require attribution, restrict use above a size threshold, or reserve rights for the publisher.

Research-only licences prohibit commercial use entirely. Easy to miss and enforced.

Derivative work terms. Some licences extend to models fine-tuned from the original, and to their outputs.

Output ownership is addressed inconsistently and sometimes not at all.

What to check before building on one: may I use it commercially, are there prohibited uses relevant to my product, does the licence bind fine-tuned derivatives, do I need to attribute, and does anything change at scale. Five minutes, and the alternative is discovering a restriction after the product exists.