Explainer · 07
Where do the models come from?
The AI on a private machine isn't something cobbled together in a garage. It's an open-weight model — built by one of the large AI labs and released publicly, so anyone can download it and run it on their own hardware.
What open-weight means
Underneath, a model is a very large file of learned numbers — its weights. Most cloud AI keeps those weights secret, so the only way to use it is through the company's servers.
Open-weight models publish that file. Download it once, and it runs entirely on your computer, without checking in with anyone each time you use it.
Who makes them
-
Meta
The Llama family, one of the most widely used open model lines.
-
Mistral
A French lab releasing capable models in a range of sizes.
-
Google
The Gemma family, open relatives of its Gemini models.
-
Alibaba, DeepSeek and others
Qwen, DeepSeek and a fast-moving field of alternatives, many of them particularly strong at coding and reasoning.
How good are they?
The best open models sit close behind the best cloud models. For everyday writing, summarising and coding, the difference is often hard to notice; for the hardest reasoning, the cloud frontier is still ahead.
That gap has been narrowing. And because models are just files, the machine you own today can often run a better model next year on the same hardware.
How one gets chosen for you
- Fit. The model has to fit in the machine's graphics memory to run well.
- Work. Some models are stronger at code, others at writing or at working through long documents.
- License. Each open model comes with its own license terms, and they're checked against how you plan to use it.