gpt-oss is fp4 - they're saying they'll next try mid size one, I'm guessing gpt-oss-20b then large one, i'm guessing gpt-oss-120b as their hardware is fp4 friendly

I think they are talking about the transistors that apply the weights to the inputs.