[stekiclaw]

Problem and result

This adds community measurements for Flash-Next IQ3_S on an NVIDIA GeForce RTX 3070 8 GiB, Intel Core i9-10900K and 61.3 GiB RAM, using an existing 65,536-token serving configuration.

The unmodified community speed suite produced 33 measured request records with no recorded request errors and no skipped default tests. Median results were:

• Short-chat decode: 46.0 tok/s; TTFT: 0.676 s.
• Fresh 4K prompt processing: 411.2 tok/s; TTFT: 9.802 s.
• Fresh 16K prompt processing: 367.9 tok/s; TTFT: 43.682 s.
• Agent follow-up conversation-cache hit rate: 71.102%.

This is the quick community suite described in docs/COMMUNITY_BENCHMARKS.md, not the thinking-enabled draft standard benchmark. It does not claim a performance improvement over another configuration.

Related discussion

None directly applicable.

Open and closed issues and pull requests were searched for RTX 3070 reports and contributions by this account. This is a new hardware/configuration measurement, not a resubmission of an existing report.

No feature, refactor, default change or engine modification is proposed.

Approach

Run the installed, unmodified tools/strata_bench.py against the existing loopback server, with thinking disabled and three repetitions.

The existing serving configuration was preserved: IQ3_S, 65,536-token context, int8 KV, 32,768-token KV residency, one parallel slot and MTP disabled in the profile.

The contribution is limited to:

bench/results/2026-10-28-community-rtx-3070-qwen3-8-flash-next-iq3-s/

It includes the human report, original generated README, unchanged per-request results, sanitized configuration, actual benchmark console output, partial resource samples, provenance and artifact checksums. No model weights, prepared packs or engine code are included.

The report was prepared with AI assistance from actual hardware measurements.

Validation

Passed:

• Executed the following command on the inference host, with private paths replaced by placeholders:

  /python/bin/python3.12 -u /tools/strata_bench.py --config --url http://127.0.0.1:8080 --runs 3 --out

• Runtime: Strata CLI 0.2.7, engine 0.1.52, installed runtime COMMIT a42cc3d.

• Environment: RTX 3070 8 GiB, i9-10900K, 61.3 GiB RAM, Linux kernel 7.0.0-38-generic, NVIDIA driver 595.99.02.

• Verified 33 measured records: 9 short-chat requests, 6 fresh long-prompt requests and 18 agent-session requests. All returned tokens without a recorded request error; no default tests were skipped.

• Confirmed zero reported reused tokens in the measured short-chat, 4K and 16K requests.

• Parsed the JSON artifacts, checked request counts and reported statistics, verified byte equality of the published results.json with the original runner output, and verified the artifact checksums.

• Ran git diff --cached --check and privacy-pattern checks on the publication files.

• Confirmed the serving profile hash was unchanged. Post-run service checks showed the model loaded at 65,536 tokens, idle, with no queued requests.

Failed:

• The first auxiliary resource sampler failed to start because of command-string quoting. It was repaired without restarting the benchmark.
• The replacement sampler recorded 53 samples over approximately 105.3 seconds near the end of the run. These measurements are explicitly labeled partial coverage, not whole-suite peaks.
• No inference request errors were recorded by the community runner.

Not run:

• Optional 32K prompt and concurrent-client tests; neither option was selected.
• Long-context needle tests, full-window capacity testing, broad accuracy/coding evaluations and baseline-versus-tuned comparisons.
• CUDA, HIP, SYCL and Metal builds: this is a results-only contribution with no engine or source-code changes.

Limits

• This is a speed and conversation-cache suite, not an answer-quality evaluation or a maintainer-approved standard result.
• The legacy runner defines a successful row as returned tokens with no recorded error. It does not retain terminal finish reasons or explicitly require [DONE], so the report does not claim independently verified natural-stop completion for all responses.
• Full request bodies and generated replies were not persisted by this runner. Byte-identical replay of the agent conversations is not claimed.
• The server was already loaded and its expert cache was warm. These are not cold-model measurements.
• Actual fresh long inputs were 4,019 and 16,058 tokens. A configured 65,536-token context is not evidence of a full-window capacity or recall pass.
• Output caps were 100 tokens for short chat, 24 for long-prompt tests and 64 per agent turn. Reaching a cap does not establish answer completeness.
• Some agent requests used accepted speculative drafts. Their higher decode rates should not be interpreted as sustained general-purpose chat throughput or evidence of MTP acceleration.
• Resource monitoring covers only the final portion of the suite.
• Exact model/source/tokenizer revisions and pack-preparation provenance remain unknown and are disclosed as such.
• The original community runner did not identify the Terminal bundle's CLI version or COMMIT. These were recorded separately without modifying the original result fields.
• The benchmark used an existing server, so the community runner did not produce an engine startup log. Native per-request timing fields are retained in results.json; the attached console output is not presented as an engine startup log.