Open weight vs closed AI models, and what you actually get
Key takeaways
- Open weight means you can download the trained model, but it does not mean the training data or code are public, so it is not the same as open source.
- Closed models are paid per token through an API and your prompts leave your machine; open weight models cost hardware and effort but can run privately.
- Reflection AI, backed by Nvidia, is preparing a first open weight model aimed at Western enterprises as an alternative to Chinese options like DeepSeek.
This article contains affiliate links. We may earn a small commission if you make a purchase, at no extra cost to you.
Last updated: 5 October 2026
Reflection AI, a startup backed by Nvidia, is preparing its first open weight model, aimed at Western enterprises that want an alternative to Chinese options such as DeepSeek. The phrase open weight is easy to misread in that sentence.
It is often used as if it meant open source. It does not, and the gap decides what you can actually do with a model.
What weights are
A trained model is, at heart, a very large file of numbers called weights. They are the result of training, and they are what make the model answer the way it does. Open weight means the lab publishes that file so you can download it and run it yourself.
Open source would mean more: the training code, often the data, and a licence that lets you reuse all of it. Many open weight releases share the numbers and keep the rest private, and some add licence limits on commercial use. Read the licence before you build on one.
Closed models and the API
A closed model stays on the lab's servers. You send a prompt through an API, the lab runs it and you pay per token. You get the best models available and no hardware to manage, but your data travels to someone else's systems and the lab can change or retire the model.
Open weight vs closed AI models on cost and privacy
Open weight models cost nothing to download, but you pay for the machine. Very roughly, at 8 bit precision, one billion parameters needs about one gigabyte of memory, so a model with tens of billions of parameters wants tens of gigabytes. Our guide to running a local LLM on a Mac walks through what fits on common hardware.
Privacy is the clearest win. A model running on your own machine keeps your documents there. If you want to try it, a Mac mini is a quiet and sensible starting point; the Apple Mac mini with M4 on Amazon shares one pool of memory between CPU and GPU, which helps with model size.
Closed models win on raw capability and convenience, at least for now. The frontier labs still keep their strongest systems behind an API.
Why Nvidia would back a free model
Nvidia sells the chips that every model runs on, so more people running more models helps it whichever lab wins. That is our reading, and it fits the broader pattern in our coverage of Nvidia's AI factories and their power costs.
What to watch
The thing to check on any open weight release is the licence and the size of the smallest useful version. If it runs on hardware you already own and the terms allow your use, it is open in the way that matters to you.
