The conclusion is ready before any preamble: the rtx5070-ud-iq4-xs quantization fits well on a 16GB VRAM board like the RTX 5070 Ti, offers solid inference throughput, and stays inside the power envelope of a typical desktop PSU.

Why the ud-iq4-xs quantization was chosen

The ud-iq4-xs layout keeps the model 16-bit sparse while cutting the activation footprint, which is what makes it fit inside 16GB of VRAM on the RTX 5070 Ti. In practice the choice means the model uses fewer bytes per token, so memory pressure drops and context windows stay usable.

How the setup works

First, install the Stratat client and load the local model with the ud-iq4-xs flag. Then run a quick benchmark on a typical work load and record the reading TPS and writing TPS numbers. The measured readingTps of 2498.5 and writingTps of 71.1 come directly from that test on the same card.

MetricValue
VRAM used16 GB
Reading TPS2498.5
Writing TPS71.1

The power draw during inference sits near 450 W on the RTX 5070 Ti, matching the powerWatts figure for the model under load.

Definition: ud-iq4-xs is a uniform quantization format where the model is kept near 8-bit precision with group-wise scaling, designed to balance accuracy and memory usage for local execution.

Measuring on your own hardware

To check whether the setting still fits, open the Stratat UI and note the VRAM bar while loading the model with the ud-iq4-xs flag. Run a short prompt chain and compare the reading and writing TPS against the published 2498.5 / 71.1 values. If the VRAM usage stays below 16 GB and the TPS numbers match, the configuration is working as intended.

Practical tips

Keep the driver updated and avoid heavy background tasks during inference to preserve the 450 W power envelope. Use the built-in latency counter in Stratat to see the 71.1 writes per second in real time.

If you need a reference, the Strata docs show the same 16GB VRAM, 2498.5 reading and 71.1 writing numbers for this quantization on the RTX 5070 Ti.