Decode · tok/s
32.3
Prefill
—
tok/100W
13.5
TensorFold 把同一套 EXL3 2.05bpw 从 11.63 拉到 32.32 tok/s（2.78x），单机吃下 262k 上下文。单卡 GB10。
Recipe writeup on one GB10. Rank is the quoted 32.32. 11.63 is the prior path in the same sentence. Tilde long-context figures stay off the rank.
EXL3