#2493 bench: community report - RTX 5060 Ti 16 GB, Coder IQ1_M, 204800-token context (0.1.44.1)

Description

## Summary

Results-only community benchmark submission: adds `bench/results/2026-10-15-community-rtx-5060-ti-qwen3-8-flash-next-coder-iq1-m/` per `docs/COMMUNITY_BENCHMARKS.md`. No engine or source changes.

Produced by the repository's own `tools/strata_bench.py` (suite 1, 3 runs per test, medians with ranges) against the HTTP API of a Strata server started from the install configuration. Model loading is not timed; a warm-up chat and a 1K prompt are discarded first. Speed only — **no correctness check was run.**

## What changed

- `bench/results/2026-10-15-community-rtx-5060-ti-qwen3-8-flash-next-coder-iq1-m/README.md` — the report (machine-generated header, format of the docs template).
- `.../results.json` — per-request data: prompt/generated tokens, engine timings, cache reuse, errors, plus the machine facts and the launch command.
- `.../engine-log-excerpt.txt` — the engine's start-up lines from the run.

Nothing else is touched; no model files or packs are included.

## Hardware, model, settings

- **GPU:** NVIDIA GeForce RTX 5060 Ti, 16 GiB VRAM (16,311 MiB reported), driver 616.92, CUDA version not read.
- **CPU / RAM:** 13th Gen Intel Core i5-13400F (10 cores / 16 threads), 31.8 GiB installed RAM. Windows 11 (10.0.26300).
- **Strata / engine:** 0.1.44.1 (project downloaded as a release, not a git checkout, so no commit hash: `strata_commit: null`).
- **Model:** `qwen3.8-flash-next-coder-iq1_m` (Coder IQ1_M, GSQ/RCO, two GGUF shards) with `--vision` and the vision encoder loaded.
- **Context:** 204,800 tokens, parallel slots 1, KV `q4_0`, `--resident-experts`, `--expert-cache auto`, `--prefill auto`, `--spec 4`, `--mtp rt`, `--vram-reserve-mib 700`, `--pool-workers 5`, adaptive tier 160/0.97.
- Launch command verbatim (credentials removed, none present):

```text
strata.exe --pack coder-iq1_m --native Qwen3.8-Flash-Next-GSQ-RCO-IQ1_M-00001-of-00002.gguf --ple-gguf Qwen3.8-Flash-Next-GSQ-RCO-IQ1_M-00002-of-00002.gguf --expert-profile expert-profile-coder.bin --expert-cache auto --prefill auto --spec 4 --mtp rt --max-context 204800 --kv q4_0 --resident-experts --vision --vram-reserve-mib 700 --control-vector-scaled Qwen3.8-Flash-Next-experimental-speed-projection.gguf:1.0 --control-vector-layer-range 4 44 --cvec-mode project --cvec-dir per-layer --spec-min-p 0.5 --pool-workers 5 --adapt-every 1 --adapt-swaps 160 --adapt-decay 0.97
```

Note for readers: this run has the **experimental speed projection** control vector enabled and `--resident-experts` — it is not a stock configuration, so it should not be read as the model's default numbers.

## Results (medians, range in parentheses)

| Configuration | Prompt tokens | Reused | Generated | Runs | Prompt tok/s | Decode tok/s | TTFT s |
| --- | ---: | ---: | ---: | ---: | --- | --- | --- |
| short chat | 46 | 0 | 100 | 9 | not measured | 34.5 (32.8–47.8) | 0.73 (0.65–0.88) |
| 4K prompt, fresh | 4,019 | 0 | 24 cap | 3 | 1,261.1 (1,125.7–1,281.2) | not measured | 3.23 (3.16–3.62) |
| 16K prompt, fresh | 16,058 | 0 | 24 cap | 3 | 1,466.7 (1,441.1–1,515.9) | not measured | 10.98 (10.64–11.19) |
| agent session, follow-up turns | 14,533 at the last turn | 68.7 % hit | up to 64 | 3 | see README | 39.6 (29.1–59.1) | 2.70 (2.00–3.15) |

Agent-session conversation cache: 68.7 % of follow-up prompt tokens reused; the 13,728 new tokens of the appended files are read once; the first turn (nothing cached) has a 1.59 s first token. Per-turn table is in the README. Total suite time 153 s.

## Limitations (also stated in the report)

- **Speed suite only.** No needle/recall test, no HumanEval, no tool-call check — quality is entirely untested here.
- Background workloads and power limits were **not recorded**; the GPU was not isolated from the desktop.
- Storage type and PCIe link width/speed not measured; CUDA version not read.
- The tool reports `strata_commit: null` because this install is a downloaded release rather than a git checkout.
- Single machine, Windows-only; short-chat decode spread (32.8–47.8 tok/s) shows this box is on the noisy side for 100-token answers.

## Extra Notes

Report was read before sharing; paths in `results.json` are already placeholdered by the benchmark (`C:\Users\<user>\...`), and no user, host or network identifiers are present. Happy to re-run anything specific (e.g. `--long` at 32K, or needle checks) if you want it alongside this.