#2156 bench: RTX 4090 community results (flash-next UD-IQ4_XS 204K/131K, Swift 1.5 IQ3_S)

Descripción

### Community benchmark results — NVIDIA RTX 4090 (2026-10-12)

Three speed-only benchmark reports from my machine, produced with the built-in
`tools/strata_bench.py` (BENCH.bat). Results-only PR — no engine changes.

**Hardware (all runs):** NVIDIA GeForce RTX 4090 24 GiB · AMD Ryzen 9 5900X
(12c/24t) · 63.9 GiB RAM · Windows 11 (10.0.26200) · driver 617.42 · Strata 0.1.43.

**Reports added (bench/results/):**
1. `2026-10-12-community-rtx-4090-qwen3-8-flash-next-unsloth-ud-iq4-xs`
   — flash-next-unsloth **UD-IQ4_XS**, 204,800-token context, `--kv int8`,
   `--resident-budget-gib 37`. Chat decode 38.6 tok/s, 4K prompt 757.1,
   16K prompt 1465.6, agent-session decode 61.4, cache hit 68.7%.
2. `2026-10-12-community-rtx-4090-qwen3-8-flash-next-unsloth-ud-iq4-xs-3`
   — same model, **131,072-token context**, `--resident-budget-gib 38`, plus
   adaptive expert-cache flags (`--adapt-every 1 --adapt-swaps 80 --adapt-decay 0.97`).
   Chat decode 32.4, 4K 752.8, 16K 1470.0, agent decode 48.4, cache hit 68.8%.
3. `2026-10-12-community-rtx-4090-swift-1-5-iq3-s`
   — **Swift 1.5 IQ3_S** (with `--vision`, `--ple-gguf`, `--vram-reserve-mib 700`),
   131,072-token context. Chat decode 88.9, 4K 1231.7, 16K 2409.2,
   agent decode 115.8, cache hit 68.7%.

**Method:** `python tools/strata_bench.py`, 3 runs per test, medians with ranges;
thinking off, temperature 0; model loading not timed.

**Limitations:** speed-only suite — no correctness/recall check was run. Background
load and power limits not recorded. CUDA version not read. Each folder contains
`README.md`, per-request `results.json`, and an `engine-log-excerpt.txt`.