#1725 bench: community report - RTX 3060 12 GB, Flash-Next Q2_0/IQ2_XS/IQ3_XXS/IQ3_S, engine 0.1.41

本文

## Summary
Community benchmark for the official Strata 0.1.41 Windows release on one RTX 3060 12 GB (Core i7-12700, 64 GB RAM, Windows 11), covering the original Flash-Next Q2_0, IQ2_XS, IQ3_XXS and IQ3_S. Results only: no engine, build or test changes.

Median prompt and decode tok/s at 4K, 32K and 128K prompt tokens, three runs each:

| Model | Prompt | Decode |
| --- | --- | --- |
| Q2_0 | 882 / 981 / 933 | 42.9 / 41.8 / 38.6 |
| IQ2_XS | 829 / 927 / 892 | 45.5 / 40.8 / 37.6 |
| IQ3_XXS | 681 / 884 / 867 | 26.2 / 27.1 / 27.0 |
| IQ3_S | 568 / 820 / 803 | 20.2 / 20.2 / 18.9 |

## What changed
- Add `bench/results/2026-10-09-community-rtx-3060-0141/`: the README (method, tables, column notes), per-condition medians (`measurements.csv`), one row per request (`requests.csv`) and per concurrency answer (`concurrency.csv`), the arguments and environment of the 86 launches behind the tables (`launches.json`) with their engine logs (`engine-logs/`), model and runtime manifests, quality results, code samples, and the harness scripts. Local paths are replaced by `<workspace>` and `<home>`.
- Add an entry to `docs/COMMUNITY_BENCHMARKS.md`. The `bench/results/COMMUNITY.md` row is left to the maintainers.
- Not attached (about 88 MB): per-request response and prompt files, 1 Hz telemetry and thermal samples. The README lists what is missing.

## Extra Notes
- Q2_0/IQ2_XS run without a resident expert pool; IQ3 uses a 20 GiB resident expert pool. The tables compare complete configurations, not quantizations in isolation.
- GPU thermal slowdown was observed (samples not attached). IQ2_XS at 260K was stopped by the harness memory guard (not an engine OOM finding), so those cells are empty.
- Coding is HumanEval IDs 0-39 only. The only completed soak is one 30-minute IQ3_S run, shorter than the hour asked for in `docs/TEST_REQUESTS.md`.