Nine models, measured on a DGX Spark

Each one is a lane: pinned, proved before it serves, and loaded with one button. Numbers from an ASUS Ascent GX10 with 128 GB, with their method.

Model	Measured	What it is	64 GB box
Qwen3.8-27B Qwen	71.4 tok/s greedy median, one request	Text and code, a 1M-token window	Yes
Qwen3.8-Flash-Next 176B Qwen	47.9 tok/s on code, with FLASH_TIER=context	A 176B mixture of experts on one box	No
Flash FP4-Local NVIDIA	55.4 tok/s greedy median, at its default effort	The FP4 build of the flash model	No
Nemotron 3.5 Bolt NVIDIA	140 tok/s rewriting code, one request, reasoning off	A 74B that reads images and video	Yes
Nemotron 3.5 Super VL NVIDIA	69.5 tok/s rewriting code, one request, reasoning off	A 120B with vision and a 1M-token window	No
Ornith 1.5 122B Ornith AI	42.4 tok/s greedy median, one request	A 122B mixture of experts with vision	No
Kolibri-1 Aleph Alpha	48.0 tok/s one request	German and English	No
Qwen-Image 2.1 Qwen	7.5 s with its Turbo, per 1024×1024 image	Images from a prompt	No
MiniMax-H3 MiniMax	126.2 to 127.9 s with the Turbo adapter, for 4 s of 480P video	Video with sound	No