Wonder how it will fare when infering frontier models or 10T+ parameters.

For coding, in order to get "quality" results, you will need those frontier models with their whole weights.