Comment by N1ckFG

In my experience so far, separately from privacy concerns there's a specific use case where cheap local shines--when unlimited shots on /goal with a dumber model is better than limited time with a smarter model. This looks less like the assisted-coding scenario that's commonly brought up as a good local scenario, because if you're searching for a fast and accurate solution to a single blocking problem, the bare minimum for a model that can do that is a 24GB dGPU or a 64GB Mac. Instead, this looks more like a Hermes agent on an Raspberry Pi driving OpenCode on an old gaming computer with just enough RAM and VRAM to handle an MoE, churning out something overnight that would quickly exhaust the subscription plans, like a knowledge graph for a large document corpus.