the local half: github.com/justvugg/colibri pure c, streams the experts off disk. k3's checkpoint is ~1.5 tb, so size the ssd first. for qa, use its brio mode: it scores merge / request changes / close instead of writing a review. 0 tokens generated, so a slow 2.8 trillion parameter model still finishes overnight. chatgpt sign-in lives in devin → settings → account → "continue with chatgpt".

if you're still sitting in front of your laptop waiting for your agent to finish, then you're the bottleneck. for an extra $20 a month, you get a night shift engineer that ships while you sleep. here's mine: → devin pro, $20/mo. runs on its own cloud computer, browser included → sign in with chatgpt. gpt-6 sol bills to the plan you already pay for → swe-2 is free on it right now → kimi k3 (2.8 trillion params) runs locally off my ssd as the qa agent devin builds in the cloud. kimi checks it locally. i wake up to a reviewed pr. that's how a 100x engineer moves.