#3353 [Benchmark] RTX 5070 Ti + DDR4-3600: IQ2_XS, IQ3_XXS and Coder IQ1_M

说明

## Problem and result

There were no Strata results from an RTX 5070 Ti (16 GiB, Blackwell) paired with DDR4 system RAM. This PR adds three speed reports from that machine, one per model, under `bench/results/`. No code, docs or defaults change.

| Model | Date | Short chat decode tok/s | 4K prompt tok/s | 16K prompt tok/s | Agent follow-up decode tok/s | Agent cache hit |
| --- | --- | ---: | ---: | ---: | ---: | ---: |
| qwen3.8-flash-next-iq2_xs (262K context) | 2026-10-15 | 105.0 | 1591.1 | 2282.0 | 116.9 | 68.7 % |
| qwen3.8-flash-next-iq3_xxs (262K context) | 2026-10-20 | 66.8 | 971.8 | 1592.0 | 75.2 | 68.6 % |
| qwen3.8-flash-next-coder-iq1_m (131K context) | 2026-10-15 | 48.1 | 563.9 | 1033.3 | 49.1 | 68.7 % |

These are medians of 3 runs. Each folder's `README.md` has the ranges, the full engine command line and the turn-by-turn agent session.

## Related discussion

none

## Approach

Each model used the install's own configuration, unchanged, so the numbers show what a user of this machine gets by default. Thinking was off and temperature was 0. Each report folder has the same three files as the existing community reports: `README.md`, `results.json` (per-request data) and `engine-log-excerpt.txt`.

## Validation

**Passed:**
- `python tools/strata_bench.py` (suite 1, 3 runs per test) on 2026-10-15 for IQ2_XS and Coder IQ1_M, at commit 296b59d2, engine 0.1.44.1. All tests returned tokens; results are in the two 2026-10-15 folders.
- The same command on 2026-10-20 for IQ3_XXS, at commit fd2277fb, engine 0.1.44.1 (this commit only adds bench results to 296b59d2, so the engine is unchanged). All tests returned tokens; results are in the 2026-10-20 folder.
- Environment: NVIDIA GeForce RTX 5070 Ti 16 GiB, driver 616.64; AMD Ryzen 7 5700X (8 cores, 16 threads); DDR4-3600, 63.9 GiB on 2026-10-15 and 71.9 GiB on 2026-10-20, as reported by the benchmark; Windows 10 (10.0.19045).

**Failed:**
none observed

**Not run:**
- Correctness checks: this is a speed suite, so output quality is unverified.
- The concurrency test.
- A same-day comparison against another build, as asked for in `docs/COMMUNITY_BENCHMARKS.md`. The numbers have nothing to compare against from that day.
- CUDA, HIP, SYCL and Metal builds: no code changed, so there was nothing to build.

## Limits

- Background load, power limits and RAM timings were not recorded. The RAM type and speed (DDR4-3600) are my own report; the benchmark recorded only the size.
- Installed RAM changed between the two dates (63.9 → 71.9 GiB), so the IQ3_XXS run is not on exactly the same machine as the other two.
- The CUDA version was not read.
- The reports are not added to the list in `docs/COMMUNITY_BENCHMARKS.md`. I can add them if wanted.
- The prompt tests reach 16K tokens, not the 4,096/32,768/128,000 sweep some existing reports use.

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