Cool! Thanks for running these benchmarks. Always interesting.

Out of curiosity, is the benchmark using the single allocation memory allocation mode or the dynamic allocation mode? The latter trades a bit of CPU performance for much lower peak memory usage when parsing.