OpenLLaMA models up to 13B parameters have now been trained on 1T tokens:

https://github.com/openlm-research/open_llama