Unsloth studio, on a server that I built in 2024 that was, at the time, ludicrously oversized, but is now just ok-sized to run the q4 quant of qwen flash next. (128GB ram, 16GB vram).

I really want to build something less scuffed, but, yknow. So I'm making do atm.

There were a bunch of high performance computing efforts from the 1990s and on where, if you could work out that you'd need 5 years of "today's" compute, you should obviously hang on to the money for 4 years, then spend it on the correspondingly improved hardware and finish the job in a under a year. I've been considering building a box to run OLMO (one of the only models with training data "that you wouldn't have to defend in court" as far as I can tell?) and even with Sam Altman's supply chain attack, stalling for bigger hardware still seems like the right path...