70B base model benchmarks?

[BarfingLemurs]

I care about 70B 2.3-2.55bit comparisons, for use with a single 24gb. What are the default MMLU and perplexity scores when you use a 4bpw base llama model (or those model sizes most similar to GPTQ variants)?

P.S.
Is the codebase (the quantization conversion), in development?, the nature of converting these models require a lot of bandwidth and compute, so updating and sharing large models is more difficult.