Fast, cheap, good, cannot get all three of them, right? Strata now runs Qwen 3.8 Flash on a gaming PC fast enough that people ask the obvious follow-up: is it actually good? The honest answer is that we run 3-bit packs (IQ3_S, IQ3_XXS), so there is a gap to the official API. We measured it. The official Qwen 3.8 Flash (Alibaba's API) answered 31 prompts, about 15K tokens, and Strata read the same answers token by token: Ask the official API the same thing twice at temperature 0, and it disagrees with itself on 2.5% of tokens. That's the ceiling. Local adds 1.5 points on top of that with IQ3_S, 2.3 with IQ3_XXS, and 2.6 with our fastest setting. Where the official model is sure (about 70% of tokens), local picks the same token 99.9% of the time. The rest is mostly synonym-level choices, where the API wavers too. And the pelican riding a bicycle? No sweat. The real perk of local: you know exactly which model and which weights you're running. No asking for Astra and getting quietly routed to Luna. No model shrinkflation. Fast, cheap, and good, we take at least 2.5 of them.