So far model releases that take advantage of Quantization Aware Training (QAT) have been focused on 4-bit. I’m curious what could be accomplished with a larger MoE model around 120b up to 400b. Obviously the model could not approach 8/16 bit performance, but perhaps this could be a better alternative to training a ternary LLM (1.58 bit) from scratch. At these sizes you could fit the model into consumer computers running 64/128 gb RAM and perhaps it could out perform a model at about half the s...