Micro Blossom co-designs an exact MWPM decoding algorithm and a heterogeneous CPU/FPGA architecture with fine-grained vertex- and edge-level parallelism. At surface-code distance 13 and a physical error rate of 0.1%, the FPGA prototype achieves an average decoding latency of 0.8 microseconds at 62 MHz. This is a decoding-latency result, not merely a throughput result.