Open weights

Where the model file sits decides who sees your data, what happens if a provider changes its terms, and whether a system you depend on can be taken away. Downloadable weights answer all three at once: the material never leaves your hardware, the version you tested is the version you keep, and no pricing change or discontinued product can strand you. The cost is that running it becomes your problem, and the strongest models in any given month are usually still the ones you cannot download.

Training produces a set of parameters, often called weights: the numbers that encode everything the model learnt. Publishing them means placing that file where anyone can fetch it, under a licence that says what they may do with it. What is published is the trained result, not the recipe. The training data, the code that did the training and the runs that were discarded along the way usually stay with the company.

Running downloaded weights needs somewhere to run them. In practice that means either a machine with enough memory and a capable graphics processor, or rented hardware, plus software that knows how to load the file and serve requests against it. Quantisation is the common accommodation: storing the numbers at lower precision so a model fits on smaller hardware, at some cost to quality.

Licences differ and are the part most easily skipped. Some published weights carry genuinely permissive terms. Others restrict commercial use above a certain size, forbid particular applications, or require attribution. The word open in the name is a description of availability, not a promise about what the licence permits, so the licence is worth reading before a plan depends on it.

Open weights and closed weights coexist inside a single company's range, and frequently inside a single product name. A provider may publish some model families while keeping its current flagship closed and available only through an interface. The family name on the label therefore tells you very little; which specific models are downloadable is the question.

• Install a local runner, pull a published model, and have it serving on your own machine behind a standard interface, with nothing leaving the hardware.

  Ollama

• Do the same through a desktop application instead of a command line, where downloading and running published models happens through visible controls.

  LM Studio

• Read a model family's page and notice which part is downloadable: one widely used line publishes weights you self-host while its current flagship stays closed and reachable only through an interface.

  Qwen

• Self-host published weights when your own data rules stop you sending information to a provider's hosted service, so the privacy decision and the capability decision stop being the same decision.

  DeepSeek

What is the difference between open weights and open source?

Open weights means the trained model file is published and downloadable. Open source means source code is released under a licence permitting reuse and modification. Most downloadable models publish the weights while keeping training data and training code private, so they are open weights without being open source in the usual sense.

Do I need my own hardware to use open weights?

Not necessarily. You can run downloaded weights on rented hardware, or use a hosted service that serves the same published model. Owning the machine is one way to get the privacy benefit, not the only way. If nothing may leave your premises, then yes, the hardware has to be yours.

Are open-weight models less capable than closed ones?

Generally they trail the strongest closed models of the same moment, and the gap has narrowed considerably. For a great deal of everyday work the difference does not show. The honest way to decide is to test both on your own tasks rather than on published scores, because the gap varies enormously by task.

Can I use open-weight models commercially?

Often yes, but it depends entirely on the licence attached to that specific release. Some are permissive, others restrict commercial use above a size threshold, forbid certain applications, or require attribution. Read the licence for the exact model and version before building anything that would be expensive to unwind.

• Open weights means the trained model file is published for anyone to download and run.
• It is not the same as open source: training data and code usually stay private.
• The benefit is control over data, versions and continuity, not a lower bill.
• Licences vary and set what you may actually do, so read the one attached to the release.
• A provider publishing some models does not mean its strongest model is downloadable.

• Ollama

  The common way to pull published weights and serve them locally.

• LM Studio

  The same thing through a desktop interface, and itself proprietary software.

• Qwen

  A family where the open-weight line and the closed flagship sit under one name.

• DeepSeek

  Published weights that let self-hosters separate the privacy question from the capability one.