[MallocVoidstar]

Benchmarks don't cover all possible use cases, for one. There's always the possibility that a model does better on every benchmark thrown at it, but for your specific use case it does worse in practice.