MD4 is a hash function; it processes input messages which are sequences of bits of arbitrary length. Processing time is roughly proportional to the input length, but with starting and finishing a hash function invocation imply some fixed overhead, partly due to the algorithm (initialization, padding...), partly to the idiosyncrasies of CPU (caches, branch prediction...).
Your first line says that OpenSSL computed a complete hash over a 16-byte messages, and did it again, and again... up to a grand total of 9063888 times within a time frame of 3.0 seconds. This implies a hashing bandwidth of about 48.34 MB/s (that's 16 times 9063888, divided by 3.0). The last line, on the other hand, says what happens when OpenSSL hashes (repeatedly) 8192-byte messages: it can do so 275679 times in 3.0 seconds, for a total hash bandwidth of 752.8 MB/s, more than 15 times more. This illustrates that the overhead of starting / finishing the hash function invocation dominates computation cost when dealing with small messages. The last line is, on the other hand, close to the bandwidth which can be achieved when processing a very long input.
Note that benchmarks mostly measure themselves. If you were to actually hash a lot of input bytes, e.g. one or two very big files, you would have to take into account the time needed to move the data around, in particular reading it from disk or network. The OpenSSL benchmark, by construction, plays entirely within the CPU L1 cache, making it often unrealistic.