BackendROCm
Decode67.39 tok/s
Prompt eval2148.7 tok/s
VRAM12 GB
Memory bandwidth~384 GB/s
ModelSpark-X2.5 4B
Parameters4.1B
QuantizationQ8_0
Context206K
KV quantf16
OffloadPartial (some layers in RAM)
About this result
Measured -p 512 -n 128 -r 3, per-test isolated telemetry (benchscope). Prompt-eval king (4.5x BF16) but high run variance (+/-215 tok/s); decode trails Q4_K_M. Full weights: 4.07 GiB.
Status: Measured in the VelsTech Lab on real hardware. Read the full Lab report →