#2173 Community benchmark: RTX 5070 Ti 16 GB, Threadripper 7960X, 64 GB RAM (IQ2_XS, IQ3_XXS, IQ3_S)

说明

Results only, no code changes. Three folders under `bench/results/`, one per model, each written by `tools/strata_bench.py --long`.

**Machine:** RTX 5070 Ti 16 GB (driver 617.42, 300 W limit), Threadripper 7960X (24 cores, AVX-512), 64 GB RAM, Windows 11 26220, model files on a Crucial P310 NVMe SSD.

**Software:** official Windows release engine 0.1.43.1, CUDA 13.0, commit a0116b27. Original Qwen3.8-Flash-Next, 65,536-token context, 8-bit KV with a 32,768-token resident window. Settings are the ones setup wrote, plus `--vram-reserve-mib 1125` added by hand. Engine 0.1.38 had reported 87 MiB of VRAM free with IQ2_XS loaded and suggested that value.

**Results** (medians of 3 runs; the chat row is 9 requests, the agent row 15 turns):

| | IQ2_XS | IQ3_XXS | IQ3_S |
| --- | ---: | ---: | ---: |
| Experts in VRAM at start | 6,477 | 5,017 | 3,845 |
| Short chat decode, tok/s | 123.6 | 115.0 | 96.3 |
| 4K prompt, tok/s | 2,795 | 2,995 | 2,865 |
| 16K prompt, tok/s | 3,287 | 3,473 | 3,448 |
| 32K prompt, tok/s | 3,320 | 3,474 | 3,458 |
| 32K time to first token, s | 9.70 | 9.27 | 9.32 |
| Agent session decode, tok/s | 130.0 | 126.3 | 102.2 |
| Agent cache hit, % | 68.6 | 68.8 | 68.8 |

Ranges, per-request data and engine start-up lines are in each folder.

**Observations and limits**

- IQ3_XXS read prompts slightly faster than IQ2_XS at every length here. That is what was measured; I did not look into why.
- Short-chat decode ranges are wide (about ±15%) because the three questions differ, so draft acceptance differs.
- Speed only: no needle or other correctness check was run with this engine.
- A normal desktop session stayed open (VS Code, PowerToys, WhatsApp, Defender), with no game or other GPU work. Power plan Balanced. The PCIe link under load was not measured.
- Each model's suite ran once, back to back with the others, each in a fresh server process. There is no repeat on another day.

🤖 Generated with [Claude Code](https://claude.com/claude-code)