Best local AI model

There's no single best local AI model for everyone: the best one is the most capable model that fits in your graphics card's memory and does the specific task you need well. A brilliant model that doesn't fit on your machine will be painfully slow or won't even start, and a well-chosen small one can be more than enough for summarising texts or sorting emails.

The rankings doing the rounds online help little, because they're measured with tests that look nothing like your work and because they change every few weeks. What's more, each family has its strengths: some write better in certain languages, others code better and others reason better but run more slowly. The sensible way to choose is the opposite of how it's usually done: first you decide the task, then you look at how much memory you have and, with those two facts, you try two or three candidates with real examples of your own.

It's the question I ask myself every time a new model appears, and the answer always depends on what I want it for. Which ones I use today (my table dates from September 2026), for what and on which machine, and the practical rule I follow to choose based on memory and task, I explain in "More than fifty models tested at home". What they're really good for day to day, and what they aren't, is in "What is having artificial intelligence at home actually good for?".