#2680 Community bench: RTX 4080 SUPER / 7950X3D – Flash-Next IQ3_XXS 128K and OrcaRouter Uncensored IQ3_XXS 64K

说明

Two community benchmark folders for a single **RTX 4080 SUPER 16 GB** (Ryzen 9 7950X3D, 63.7 GB RAM, Windows 11), each with README, per-request `results.json` and `needles.json`.

| Folder | Model | Key settings | Short-chat decode | 4K / 16K prompt | Agent follow-up | Recall |
|---|---|---|---|---|---|---|
| [flash-next-iq3xxs-128k](bench/results/2026-10-15-community-rtx-4080-super-7950x3d-flash-next-iq3xxs-128k/) | Qwen Flash-Next IQ3_XXS, 131K | `--pcie-frac 0.35 --spec-min-p 0.70 --adapt-every 1 --adapt-swaps 80 --adapt-decay 0.97`, int8 KV, MTP | 94.8 tok/s | 2,396 / 3,297 | 104.4 tok/s | 6/6 @ 32K & 128K |
| [orcarouter-uncensored-iq3xxs-64k](bench/results/2026-10-15-community-rtx-4080-super-7950x3d-orcarouter-uncensored-iq3xxs-64k/) | OrcaRouter Flash-Next Uncensored IQ3_XXS, 64K | calibrated `--pcie-frac 0.20 --spec-min-p 0.70 --adapt 1/160/0.97 --adapt-async 1 --vram-reserve-mib 1000`, `STRATA_PCIE_MIN1` | 95.4 tok/s | 2,017 / 3,517 | 109.7 tok/s | 6/6 @ 32K & 64K |

Method: `tools/strata_bench.py` suite 1 (3 runs per test, medians with ranges), engine 0.1.44.1 (commit 296b59d2). Context recall via `tools/needle_bench.py`, six needle-in-a-haystack checks per model, **6/6 found** at full depth (125K for the 128K ctx model, 64K for the 64K ctx model). No correctness suite (e.g. HumanEval) run. Both rows also added to `bench/results/COMMUNITY.md`.