thomas

[Sudo su @sudoingX]

the honest catch, there's always one. 1bit is not full precision, you feel it on the hardest single-shot reasoning, the kind of problem you'd hand a frontier model anyway.

but automation isn't that. automation is a hundred small structured steps in a loop, and 1bit holds for structured work and tool calls fine. for 200 dollars and 8 gigs running unattended, that trade is absurdly in your favor.

speed isn't the point but for the record, 42 tokens a second fresh, about 21 deep into a long generation. not a chatbot race, a background worker. it doesn't need to be fast, it needs to not stop. it doesn't.

the floor for local automation just dropped to a card you might already own. not a rented gpu, not a cloud key, the one in your drawer plus a 3.5 gig download.