[As7656431]

Problem and result

A results-only community report: an RTX 5070 12 GB with 31.8 GiB of RAM on Windows 10, the original Flash-Next IQ2_XS in low-RAM mode (--mmap-experts), Strata 0.2.3.1, measured with tools/strata_bench.py. As far as I can see there is still no single-GPU RTX 5070 (non-Ti) report in bench/results/, and this PC has a slower NVMe drive than the 31.6 GiB RTX 5070 Ti report in #2791.

On 0.2.3.1: 4K prompt 87.0 tok/s (first token 46.3 s), 16K prompt 101.3 tok/s (158.7 s), agent follow-up turns 62.8 s to the first token and 5.7 tok/s decode, short-chat decode 55.8 tok/s.

Related discussion

none required for a results-only report. This PR was closed in the 0.2.0-beta reset; it held the same PC's report for 0.1.45, which this one replaces. The opt-in changes for this setup are discussed in #3065 (#3072, #3086).

Approach

• bench/results/2026-10-23-community-rtx-5070-qwen3-8-flash-next-iq2-xs/: README.md, results.json and engine-log-excerpt.txt, new.
• The 0.1.45 folder bench/results/2026-10-18-community-rtx-5070-qwen3-8-flash-next-iq2-xs/ is removed from this PR.
• No engine or code changes. docs/COMMUNITY_BENCHMARKS.md is left alone so you can decide where to list it.

Validation

Passed:

• PYTHONUTF8=1 python tools/strata_bench.py --url http://127.0.0.1:8080 --config <the config> on 2026-10-23, the tool of 7ddc17a (v0.2.3.1) against serve/server.py of the same commit; the engine was built from that commit on this PC (CUDA 13.0, sm_120). The suite finished in 2161 s with every request answered; the tool wrote README.md and results.json.
• The README was then completed by hand: memory, drive, PCIe and locale, how the engine was built, the one setting left out (the optional control vector), how the tool was run, and a Notes section. The numbers are the tool's.
• engine-log-excerpt.txt is the benchmark run's part of the engine log, cut with the tool's own log_excerpt() and redactor(); local folders are replaced with placeholders. All three files were checked for user names, host names and local paths.

Failed: none observed.

Not run: the release package's own engine binary (built from the release commit instead), the standard benchmark protocol (strata bench run, still a draft), --long (32K context), --clients, a RAM budget, other models.

Limits

• Speed only, one machine, one model, 32K context.
• Two ways of running the tool on this PC: the server was started separately (so the tool did not save the engine log itself; see engine-log-excerpt.txt), and PYTHONUTF8=1 because on a Chinese-locale Windows the tool still reads CMakeLists.txt with the default code page (tools/strata_bench.py: UnicodeDecodeError reading CMakeLists.txt on non-UTF-8 Windows (GBK/cp932), before any measurement #2792).
• The daily configuration was used except that the optional control-vector file was left out, so the numbers compare with other reports.

[As7656431]

Related: #3065 describes opt-in changes for this setup, with measurements.

[Niko1221]

Thanks for the pull request.

We're resetting the tracker for 0.2.0-beta. There are more than 1,300 open issues and pull requests, too many for us to review properly, so we're closing everything that isn't being worked on right now. Closed here does not mean rejected.

To reopen it takes about two minutes:

1. Rebase on main if it no longer applies, and re-run your checks on 0.2.0-beta.
2. Start the title with its area: [Engine], [GUI], [TUI], [Benchmark], [Docs] or [Other].
3. Fill in the pull request template: the problem and result, and the validation (what you ran, on which revision, what you saw).
4. Comment /reopen. A check runs and either reopens the pull request or lists what is still missing.

The rules are in CONTRIBUTING.md.

[As7656431]

/reopen

[github-actions]

Reopened, thank you. A maintainer will pick it up from here. If it conflicts with main, rebase it first; that speeds up review.

[Niko1221]

Thank you for measuring this and writing it up. We are building a Strata website with a page for community benchmarks, and results like these will be added there when it launches, so we are closing this pull request here; the numbers are not lost. If your numbers change with a new version, you are welcome to send them again.