Zember — a local AI sized to your graphics card

Pick by your card

Every package holds the largest brain that fits entirely on that card. Nothing spills into system memory, so the speed on the box does not depend on what RAM you happen to own.

If your card is small but your RAM is not

These two are different: the brain is bigger than any card we list, so most of it sits in system memory. That means the speed is set by your RAM, not your card — a 6 GB card with fast memory beats a much more expensive card with slow memory. Nobody else prints this, because nobody else measured it.

Why the numbers have labels

ESTIMATED is on every speed here, and it is honest: we do not own most of these cards. The figure comes from the card's memory bandwidth and an efficiency number we measured on five brains with the very engine in the box — not one borrowed from a chart.

BRAIN TESTED means we started that exact brain and made it write, on the card we own (an RTX 5070, 12 GB). For the small packages that card is faster than yours, so it is evidence of method, not a promise. For the big ones part of the brain sat in system RAM, so on the card they are made for they go faster. MEASURED is the speed we got on that machine.

Contract test: a 3-page contract built to trip it up (the real late-payment penalty in the last clause, wrong figures on the way, a later clause replacing an earlier one), four questions, each asked three times, through the very app in the box.

Every package carries a MEASURED.txt with its figures and how each one was produced. Each package is a separate product with its own key.