They don't, they're playing "hide the #s" a bit. Llama 3.2 3B is definitively worse than Phi-3 from May, both on any given metric and in an hour of playing with

They don't, they're playing "hide the #s" a bit. Llama 3.2 3B is definitively worse than Phi-3 from May, both on any given metric and in an hour of playing with the 2, trying to justify moving to Llama 3.2 at 3B, given I'm adding Llama 3.2 at 1B.