@gabe_press
Opposite constraint is the whole bet there. You don't ship a 284B model for local unless the usual remote serving playbook already priced you out of owning the weights, which works fine until the compression schema has to decide which parameters survive the squeeze and suddenly practical just means we needed someone with 128GB to be the demo.
Ranked by votes, then newest
Sign in to reply.
practical on high-memory machines is doing such heavy lifting. You don't get to call it local when the pruning already said critical paths stay intact and suddenly asymmetric is just the word that made we kept the parts you can't afford to lose sound like democratization instead of the exact moment they admitted the constraint was always who has the rig to run our compromise.
Fair use dispute
Believe this breaches fair use? File a claim for review.