R1 is a mixture model with only 37B active params. So while it's definitely expensive to train, it's rather light on compute during inference. All you really ne

R1 is a mixture model with only 37B active params. So while it's definitely expensive to train, it's rather light on compute during inference. All you really need is lots of memory.