#1972 bench: community report - RTX 3070 8 GB, Ryzen 9 5900X, 96 GB: agent-coding runs, quantization and thinking effort, sampling A/B for #1274

说明

## Title
bench: community report - RTX 3070 8 GB, Ryzen 9 5900X, 96 GB: agent-coding runs, quantization and thinking effort, sampling A/B for #1274

Issue: refs #1274 (the A/B its author asked for); no issue to close.

## Summary
A community benchmark report from a single RTX 3070 8 GB (Strata 0.1.41, Flash-Next IQ2_XS / IQ3_S / IQ3_XXS, OpenCode 1.18.16):

1. six local configurations on the same eight agent-coding tasks (all passed 8/8; total time from 68 min 32 s to 107 min 54 s; decode 30.8 to 46.5 tok/s),
2. the same tasks with three paid hosted agents, labelled "for reference only" (same precedent as the 2x5060ti report),
3. the greedy-against-thinking-preset A/B that the author of #1274 asked for: with the thinking preset, decode speed was 0.94 of greedy (median of 57 same-prompt pairs), total wall time +12.2%, draft acceptance 72.7% against greedy's 81.2%; exact answers 25/30 against greedy's 26/30, tool calls valid 15/15 in both, no looped answer in either arm (0/57).

## What changed
- New `bench/results/2026-10-11-community-rtx-3070-8gb/README.md` with hardware, software, model, settings, workload, every number, and what was not tested.
- New `bench/results/2026-10-11-community-rtx-3070-8gb/data/`: `agent_runs_summary.json` (numbers only), `ab1274.py` (the A/B script; synthetic prompts), `ab_results.jsonl`, `ab_RESULTS.md`.
- One list line in `docs/COMMUNITY_BENCHMARKS.md` ("Community reports"):

  ```
  - [2026-10-11: RTX 3070 8 GB, Ryzen 9 5900X, 96 GB RAM](../bench/results/2026-10-11-community-rtx-3070-8gb/README.md):
    Strata 0.1.41, original Flash-Next IQ2_XS / IQ3_S / IQ3_XXS, 196,608-token context (the sampling A/B at 262,144); agent-coding runs with OpenCode
    (eight tasks, six configurations), a paid-agent reference table, and a greedy-against-thinking sampling A/B. No 4K/32K/128K sweep.
  ```

No engine, server or web code changes.

## Extra Notes
- The eight tasks are private; the report gives settings and numbers only, so the task results cannot be repeated by others. The A/B (section 3) can: the script and its prompts are in `data/`.
- Section 2 is a reference table with the limits written next to it (different tools and hardware, a reviewer from the same model family as one candidate, eight tasks, one run each).
- Not tested: the 4,096 / 32,768 / 128,000-token sweep, recall checks, vision, more than one GPU, engines 0.1.42 and 0.1.43 (released after these runs).
- `data/ab1274.py` is the script that ran, with three changes made afterwards (no private defaults, no waiting when the server has no `/queue`, a report with generation time, totals and a repeatability count); the code that sends requests and scores answers is unchanged. `python3 data/ab1274.py selftest` runs it against a fake server.
- Checked before opening: every number in the README against the harness outputs; no keys, hostnames, paths of private projects or task text in any file.

Prepared with the help of Claude Code.

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