Decode · tok/s
127
Prefill
—
tok/100W
34.3
New machine. RTX 5060 Ti 16GB plus an RTX 3070 8GB, only 32GB of RAM. Strata 0.1.44 running Qwen3.8 Flash Next Coder at IQ1_M: 127 tok/s decode with 154K of context loaded. Peak 154. Whole rig pulling 219 W. MTP is accepting 97% of drafts at 3.72 tokens per step.
First-person dual-card rig. Rank is the quoted 127 decode at 154K. Peak 154 stays in the note. Model is the Coder prune. No prefill number in the quote.
