VDAOvideo

Is Frontier Class Local AI Finally Practical?

Codacus · 286K views · 2 months ago
Description

A 177-billion-parameter model, running on a five-year-old RTX 3060 with 12 GB of VRAM. Level with Claude Opus 4.8, going toe to toe with Opus 5 in a lab built specifically to trip it up — and not crawling. The performance isn't even the crazy part. The crazy part is the architecture: a second "phrase book" of 51 billion parameters that isn't a neural network at all, it's a lookup table — and it can live on the SSD instead of the GPU. No open-weights lab has shipped this at frontier scale before. In this video I explain how it works, get it running on real consumer hardware, benchmark every RAM size from 12 GB to 61 GB, and then put it through a three-task coding lab built to lie to the model — with tests that contradict the spec, decoy legacy files, and hidden checks the model never sees