What's this about?
A detailed forum post arguing that people running models on their own hardware are usually not seeing what the model can actually do, because of configuration problems rather than the model's real ability.
What exactly is "quantisation"?
Running a model locally usually means shrinking it — storing each of its numbers at lower precision so it fits in your memory. Do it lightly and quality barely moves; do it aggressively and the model gets noticeably worse. Combined with wrong prompt templates and truncated context settings, this is why a local model that should match a hosted one often does not.
What's the community saying?
The thread is unusually practical for HN, focused on specific settings — quantisation levels, correct chat templates, context length limits — and largely agrees the gap between a well-configured and badly-configured local setup is bigger than the gap between models. It pairs directly with the Apple 512GB story, since more memory means less aggressive shrinking.