Open-Weight vs Closed Models
Level 4 · Choosing and running a model
The difference between a model you download and one you reach online. What a frontier model and a dated snapshot mean when you pick one.
Course contentsLesson 11 of 15
Level 1 · What an LLM is
Level 2 · Inside the model
Level 3 · How a model is trained
Level 4 · Choosing and running a model
Level 5 · Where LLMs go wrong
After this lesson you can
- Choose between open-weight and closed
- Say what a frontier model is
- Pin a snapshot so behaviour stays the same
12 mincovers 5 glossary entries
Worth reading first: Large Language Model Basics
Open-weight or closed: what you get with each
An open-weight model is one whose weights you can download and run yourself. A closed model stays on its maker's servers, and you reach it over the internet through an app or a paid service. Llama, Mistral and DeepSeek release weights. The GPT, Claude and Gemini models are closed.
Open-weight gives you control. The model runs on your own machine, so your data never leaves it, and it keeps working offline. Nobody can change it or switch it off under you. You pay for the hardware instead of per token, which is cheaper when you use a lot and dearer when you use a little.
Closed gives you the strongest models with no setup. The maker runs the chips, keeps the model current and adds new features. You pay per token, and your data goes to their servers under their terms. You also live with their changes.
Open-weight is not the same as open source. Most releases share the weights but not the training data or the full recipe. Many come with a licence that limits use. The Llama 2 licence, for one, asked companies with over 700 million monthly users to get a separate deal.
A frontier model is one of the most capable models of its day. The label moves: this year's frontier is next year's middle. Frontier models are mostly closed, and they cost the most per token. For many everyday jobs, a model a step below the frontier does just as well.
A multimodal model takes in more than text. Most frontier models now read images, and some take in sound or video too. You can hand one a photo of a receipt or a chart and ask about it. Many still only write text back, while some can also make pictures or speech.
Models change over time, so makers name fixed versions. A model snapshot is one frozen version, often with a date in its name. A short name, by contrast, may move to a newer version without warning. For anything you test and ship, call the dated snapshot so its behaviour stays the same.
Snapshots do not last forever. Makers retire old ones on a schedule they publish in advance. When that date comes, your app has to move to a newer version and be tested again. Plan for that move on the day you pick a model, not the week it is switched off.
Which one to pick
- 1
Pick a closed model
When you need the strongest answers now. Your use is small, and your data may go to another firm.
- 2
Pick an open-weight model
When the data must stay with you, the app must work offline, or your use is large and steady.
Worked example
Jack keeps patient notes in the building
Jack's clinic wants a model to sum up patient notes, and the notes may not leave the building. A closed model is ruled out on day one.
He runs an open-weight model on one server in the back office. It has 8 billion parameters. No note goes over the internet.
The words for this lesson5
Check yourself
Can you run a closed model on your own computer?
Finished this lesson?
Marking it complete also marks its 5 glossary entries learned, which updates your roadmap and your saved page. 0 are already marked.