#2993 bench: RTX 4070 / Ryzen 7 9800X3D Windows IQ3_S community report

Beschreibung

## Summary
Community speed report for native Windows Strata 0.1.45 (477ac13c), Qwen3.8-Flash-Next IQ3_S, RTX 4070 12 GiB, Ryzen 7 9800X3D and 93.7 GiB RAM.

## Results
Three repetitions per test; medians from the supplied suite:
- Short-chat decode: 66.1 tok/s
- Agent follow-up decode: 69.7 tok/s
- Fresh 4K prompt: 1825.2 tok/s
- Fresh 16K prompt: 1794.1 tok/s
- All 33 measured requests succeeded.

## What changed
Adds only README.md and original per-request results.json under bench/results/2026-10-17-community-rtx-4070-qwen3-8-flash-next-iq3-s/. No engine, installer or default-setting changes.

## Configuration and limits
PCIe share0.35, spec-min-p0.70, MTP, adaptive expert cache and experimental speed projection enabled; exact launch flags are recorded. No heavy background workloads and default GPU power limits, as reported by the contributor.

This is one saved benchmark suite, not a randomized A/B comparison or quality evaluation. Thinking off, temperature0; 100-token chat, 24-token prompt and 64-token agent caps. Context262144 is configured; the largest measured fresh prompt was16058 tokens. No correctness, vision, concurrency or long-context recall checks. Startup not timed; no startup excerpt because the server was already running. Raw JSON is preserved.