StudioRunning Qwen3.8-Flash-Next IQ3_S on an RTX 3070 with Strata: 1000 tps prefill, 40 tps decode
Original title: Strata: Qwen3.8-FN IQ3_S on a rtx 3070 - 1000pp - 40tps decode
A user ran Qwen3.8-Flash-Next IQ3_S with Strata on an RTX 3070 plus Ryzen 5 5600 and 16GB RAM, reaching about 1000 tps prefill and 40 tps decode at 262144 context. The biggest gains over the vanilla config came from STRATA_HC_Q8 (quantizing hyper-connection matrices to free VRAM for experts) and CPU-shared prefill; a GPU unlock tweak added only 50-100 tps of prefill. They also swapped in a q8 n-gram table instead of the bundled iq4_nl one and shared the full JSON config, asking for further speedup tips.
The author shares a complete config and the key tuning changes, so readers can reproduce large-model local inference on modest hardware.
The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.