Quen 3.8 27B. A serious local coding
model. But what will it run on? Dense 27
billion parameters, 4-bit quantization,
about 18 GB. But loading the weights is
not the whole story. 16 gigs, not a Q4.
You would need aggressive quantization
or CPU offload. 24 gigs is where it gets
interesting. Q4 fits a single 24 gig
GPU. A 24 gig Mac [music] squeezes it
in, but that memory is shared with Mac
OS. It fits. Not the tier I would buy.
32 gigs of unified memory.
>> [music]
>> Now it makes real sense on a Mac. 48 to
64, real breathing room for context and
agents. And a 32 gig GPU keeps the
entire model accelerated. So yes, a
serious 27 billion parameter model runs
on consumer hardware. On a Mac, 32 gigs,
64 to use it like an agent. Quen 3.8 is
only the beginning. Comment the model
you want me to test next.