You can estimate the model size by looking at tokens per second and comparing to open source models