It is ~260GB with presumably fp16 weights. Should fit into 64GB at 3-bit quantization (~49GB).

Edit: To add to this, I've had good luck getting solid output out of mixtral 8x7b at 3-bit, so that isn't small enough to completely kill the model's quality.