Aren't they doing the exact same thing as RedPajama?
How is this not a duplicate effort? Or are they working together with the RedPajama project? If so, why use the OpenLLaMA name?
This led to differences in evals, now they are bringing in more modern chips.
Aside from some tokeniser differences this is a drop in replacement for existing LLaMA that matches performance.
I don't think 30b and 65b are useful given what we do, the key is optimising models for consumer & swarming them.
As for SD.. Maybe try the bot on the discord server testing the new version: https://discord.com/invite/stablediffusion
The jump in quality is very significant as well.
65b is definitely a lot less common.
What other differences are there? The page doesn't mention how numbers are handled. These are two major things that impact model reasoning and numeric ability.
We can now set and forget runs so will have a better dataset for the next 13b and different tokeniser, this was meant to match as close as possible to be a drop in.