Batch 1 · single stream
From hardware specs to an honest range.
Short-context decode at s = 1,024. Pick a model and machine; we show the declared sensitivity range - or explain why we cannot.
Fits9.7 GB of 80.0 GB usable
Estimated range · llama-cpp
138.5–189.2
tokens/s at s = 1,024
Weights-only ceiling234.4 tok/s
Assumption: the selected backend target's declared bandwidth-efficiency interval. Not measured. Fit is evaluated at s = 1,024 with an fp16 cache.