GLM-130b is available and uses 8-bit quantization, was trained on 400 billion tokens, and runs on 4x3090, which, of course, is why all the cheap Vast.AI instan

GLM-130b is available and uses 8-bit quantization, was trained on 400 billion tokens, and runs on 4x3090, which, of course, is why all the cheap Vast.AI instances are gone :)