Decode · tok/s
80
Prefill
—
tok/100W
33.3
nemotron 3.5 bolt (74b) running at 80 tok/s on one dgx spark 2-bit gguf + nvidia's lilicorr drafter on llama.cpp, 262k context. - math 108 tok/s - code 81 tok/s - summary 70 tok/s - prose 63 tok/s with a 131k prompt it still does 70 tok/s.
First-person one DGX Spark. Rank uses the quoted 80 tok/s headline. Breakdowns stay in the note. 2-bit GGUF + NVIDIA lilicorr drafter on llama.cpp. 262k context claimed; 70 tok/s at 131k prompt.
HF