Maybe Llama 3.3 70B doesn't count as running on "one GPU", but it certainly runs just fine on one Mac, and in my tests it's far better at holding onto concepts

Maybe Llama 3.3 70B doesn't count as running on "one GPU", but it certainly runs just fine on one Mac, and in my tests it's far better at holding onto concepts over a longer conversation than Gemma 3 is, which starts getting confused after about 4000 tokens.