#3000 bench: RTX 5090 IQ3_S report on 0.1.45

本文

## Title
bench: RTX 5090 IQ3_S report on 0.1.45

Issue: N/A

## Summary
Adds a results-only community benchmark for Qwen3.8-Flash-Next IQ3_S on a single RTX 5090 with Strata 0.1.45 (commit 477ac13c). The suite used the repository's `tools/strata_bench.py --long` with three runs per test at 4K, 16K and 32K prompts, short chat, and an agent session. No engine or configuration code is changed.

## What changed
- Added `bench/results/2026-10-18-community-rtx-5090-qwen3-8-flash-next-iq3-s/README.md`: hardware and software, model and launch settings, method, medians and ranges, agent-cache numbers, and limitations.
- Added `bench/results/2026-10-18-community-rtx-5090-qwen3-8-flash-next-iq3-s/results.json`: per-request timings and token counts for all runs.

## Extra Notes
- Tested against an already running server; model loading is not included.
- Medians: chat decode 172.6 tok/s, 4K prompt 3745.6 tok/s, 16K 6003.7 tok/s, 32K 6403.8 tok/s; agent decode 182.1 tok/s.
- Hardware: RTX 5090 (32 GiB VRAM, 600 W power limit), Ryzen 9 9950X3D, 93.3 GiB RAM, Windows 10, driver 591.86, CUDA 13.1; system SSD, read speed not measured.
- Background load was idle apart from Strata and the Claude/ChatGPT clients.
- Context is configured at 262,144 tokens; the measured prompts cover up to about 32K tokens.
- Model source is the installed IQ3_S pack; the exact upstream repository revision was not recorded.
- Speed only: no correctness, needle, or concurrency test was run.
- Single machine, single GPU and one server session; no cross-build or cross-GPU A/B comparison.