The measurements here were taken on eight H100s. Every number here carries its error bar or says why it has none — including the numbers we had to correct downward.
TP8 no speculation 183.6 → TP2 + DFlash-64 1,795.1 tok/s (9.78×, one window, n=3). In a later session on a shared node, TP2 alone ran at 112.7 tok/s and TP2 + DFlash-64 at 1,746.5: speculation alone is 15.5×. Across sessions 9.5–9.8×. Synthetic filler prompt, greedy decoding. Linear scale 9.40–10.30
latest
GLM-5.3-Flash (320B open-weight MoE) on 8× H100. Its native MTP head gives 2.04× on a single stream with a synthetic prompt and 1.52× per task in an agent loop. Without speculation, each decode step is about 1,700 small kernels, and single-stream decode runs at about 13% of the bandwidth floor.
Five hollow harness generations in one day. The fifth check, the benchmark's own gold patch, scored 0/25.