The last time I ran one locally was a few months ago, so this might be somewhat out of date, but I’ll say Qwen.

Generally there are two numbers to look at first, the number of parameters (number of nodes + much larger number of connections between nodes) and the quantisation (number of bits that define each parameter)

So an 8 billion parameter model at 8-bit quantisation will require 8GB of working RAM, probably VRAM for fast results at home (at 8 bits to the Byte)

and also a 16 billion parameter model at 4 bits quantisation will also fill the same 8GB.

in my experience, high bit depth enables the thing to respond in more complex terms, and more parameters means more “knowledge” is there to be accessed. I have better results from phrasing questions that’ll have short answers and leaning more towards more parameters than bit depth.

The next number to consider is “context window”, which is basically how far back into your chat the thing can “remember” and have previously given information influence the next answer.

A year to two ago, when I was trying a lot of LLMs for coding tips, Google’s seemed the best for larger projects, simply because it had a much bigger context window than any other, and so could “consider” longer and more interconnected scripts as a whole, even if perhaps Claude was “smarter” with a small problem. I don’t know if this is still true in autumn 2026.

They’ve (commercial models) started keeping a per-account or per-chat log of important facts that exists alongside the context window, but doesn’t carry as much detail. Good for seeming to get to know you, not as good for writing new code that won’t conflict with a long-established bit of old code.

It’s not really practical to have million+ token context windows on home/gaming hardware.

(usually a token is a word, or a couple of tokens are a long or rare word, or a token is a punctuation mark)

LM Studio is good software with generally sensible defaults for downloading pre-trained models and running them on your own PC.

It’s a one-and-done process, then the normal Q&A usage has vastly lower requirements.

Local models generally means you download the result of all that , after it’s done, (the PT in ChatGPT stands for pre-trained) and then generate answers locally, but the training wasn’t done locally. The generating answers step is called “inference”.

Hardware cost to do inference with a top-level model is within reach of most small businesses. Hardware cost to do training of a top-level model is orders of magnitude higher.

So, there are open-source models that are trained by universities, and then used by affluent but not billionaire people at home.

There are open-source models that can do inference (answer questions) on a typical 8GB gaming graphics card, but they aren’t in the same order of magnitude of quality as the famous top-level models.

training a top-level model takes harvesting petabytes / exabytes of revelant data and then months of continuous 100% power on a large server farm costing many millions of dollars.

come on, I’ve trained toy size models too, you can’t seriously say they’re even vaguely approaching the capability of models that takes a whole massive server farm weeks or months to train.

noooooo thank you we’ve had more than enough foreign actors doing bad cod-Oirish accents.

I didn’t recognise her with a facial expression.