[orangepanda]

Or maybe even middle class plebeian 24gb rigs?

[griomnib]

At that point just run 8b.

[pulse7]

Or wait for the IQ2_M quantization of 70b which you can run very fast on 24GB VRAM with context size of 4096...

[griomnib]

At some point there’s so much degradation with quantizing I think 8b is going to be better for many tasks.