I hope you continue the 50-100M parameter models.
I think there is a case for models that finish fast on CPUs in solve by llm test cases.