Most local-AI guides recommend models two generations old. Here is what fits 8, 16 and 32 GB right now.

The fastest way to waste a first weekend with local AI is to follow a guide written six months ago. Model families ship new generations every few weeks, and a tutorial that names a specific version is usually stale by the time it ranks in search.

noizz.io/local-ai takes the opposite approach: one living page that tracks the current open-source generations, with per-model RAM and VRAM numbers for Ollama and LM Studio, a quantization ladder (Q8 near-lossless down to the IQ variants), and the two gotchas that bite most people: a bigger model at Q4 usually beats a smaller one at Q8 for the same footprint, and your context window eats memory separately from the weights.

It is free to read, no signup. The rest of noizz.io is a brand discovery and AI comparison platform, which is exactly why keeping the local-AI page honest matters: it was publicly corrected once on Reddit for naming an outdated model, and the page now opens with that correction instead of hiding it.

If a number on it is wrong for your hardware, say so here. That is how it stays current.