There are larger models in there, a 8B and a 6B. By this logic they should be above 2B model, yet we don't see this. That's why we have open standard benchmarks

There are larger models in there, a 8B and a 6B. By this logic they should be above 2B model, yet we don't see this. That's why we have open standard benchmarks, to measure this directly - not hypothesize by the models' sizes or do some cross-dataset arithmetics. Also note that, Voxtral's capacity is not necessarily all devoted to speech, since it "Retains the text understanding capabilities of its language model backbone"