Qwen3.8-9B-Distill · file← Qwen3.8-9B-Distill
Q8_0
The one we kept and tested: the maker's Q8_0, run through 14 of our tests in one afternoon. Empero-ai GGUF.
- Size
- 9.8 GB
- Status
- On disk
- Scored runs
- 15
- Smallest Mac
- 18 GB estimate, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| chainwalkqwen3.8-9b-distill | 0.0% | 0.0–65.8 | Passed 0/2 | llama.cppagentic-fast | 2026-09-14 |
| contractqwen3.8-9b-distill | 25.0% | 14.9–38.8 | Passed 12/48 | llama.cppextraction | 2026-09-14 |
| enrichmentqwen3.8-9b-distill | 79.6% | 67.1–88.2 | Passed 43/54 | llama.cppextraction | 2026-09-14 |
| epistemicsqwen3.8-9b-distill | 89.6% | 77.8–95.5 | Passed 43/48 | llama.cppstudy-nothink | 2026-09-14 |
| groundingqwen3.8-9b-distill | 67.2% | 58.6–74.8 | Passed 84/125 | llama.cppextraction | 2026-09-14 |
| nullschemaqwen3.8-9b-distill | 18.1% | 10.9–28.5 | Passed 13/72 | llama.cppextraction | 2026-09-14 |
| screeningqwen3.8-9b-distill | 91.7% | 81.9–96.4 | Passed 55/60 | llama.cppextraction | 2026-09-14 |
| screeningqwen3.8-9b-distill-recheck | 91.7% | 81.9–96.4 | Passed 55/60 | llama.cppextraction | 2026-09-17 |
| synthesisqwen3.8-9b-distill | 83.3% | 72.0–90.7 | Passed 50/60 | llama.cppextraction | 2026-09-14 |
| taggerqwen3.8-9b-distill | 83.3% | 78.1–87.5 | Passed 200/240 | llama.cppextraction | 2026-09-14 |
| tagger_nullableqwen3.8-9b-distill | 90.0% | 85.6–93.2 | Passed 216/240 | llama.cppextraction | 2026-09-14 |
| toolcallqwen3.8-9b-distill | 0.0% | -0.0–6.0 | Passed 0/60 | llama.cppextraction | 2026-09-14 |
| toolsqwen3.8-9b-distill | 100.0% | 79.6–100.0 | Passed 15/15 | llama.cppagentic-fast | 2026-09-14 |
| toolscaleqwen3.8-9b-distill | 100.0% | 91.2–100.0 | Passed 40/40 | llama.cppagentic-fast | 2026-09-14 |
| worldviewqwen3.8-9b-distill | 40.0% | 26.3–55.4 | Passed 16/40 | llama.cppextraction | 2026-09-14 |
Runs marked unenforced were measured on a suite that requires constrained decoding, on a server shown not to apply the schema: mlx_lm accepts it and ignores it, and a few other servers and models break it. Those numbers do not measure what the suite is designed to measure. They are kept and labelled rather than deleted.
Rebuilt from the lab's records by Claude, an AI model. The results above come straight from the automated runs. How these pages are made.
| File | Qwen3.8-9B-Q8_0.gguf |
| Format | Q8_0, 9,786,060,096 B |
| Where from | downloaded 2026-09-14 14:17Z to ~14:25Z from empero-ai/Qwen3.8-9B-Distill-GGUF @ 760121cd |
| Status | on disk; benched, reference baseline; never seated |
| Hugging Face | https://huggingface.co/empero-ai/Qwen3.8-9B-Distill-GGUF |
Where it came from
- Repo / file / revision:
empero-ai/Qwen3.8-9B-Distill-GGUF,Qwen3.8-9B-Q8_0.gguf, commit760121cd70bb4c36b2b5ec58eb765e0df5987efe(HF download metadata on disk, read 2026-10-23). - Local path:
models/qwen3.8-9b-distill/Qwen3.8-9B-Q8_0.gguf. - sha256: the download metadata records the etag
79ca5d342a07922f2bbf38c8d892a79a3c8620c65feaf4b1c66b7830ae724db8, which for an LFS file is the content sha256; the file was not re-hashed here. - How:
uv run --quiet --with huggingface_hub --with hf_transfer python -c "...hf_hub_download('empero-ai/Qwen3.8-9B-Distill-GGUF','Qwen3.8-9B-Q8_0.gguf', local_dir=...)", started 2026-09-14 14:17:51Z. A first attempt withmax_workers=6died on an unexpected keyword. The file completed at 14:25Z (file mtime 15:25 BST). - Who asked: Alan; the assistant chose Q8_0.
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
qwen3.8-9b-distill | 2026-09-14 14:31:00Z launch; suites finished 14:31:52Z (toolcall) to 15:45:00Z (tagger_nullable); run-log ts 15:31:52 to 16:45:00 BST | .local/bin/llama-server, 0.4.0-dev (build 10893, commit c32d1dabe); manifest server.build = b10893-c32d1dabe | -c 65536 (manifest n_ctx 16384 per slot) -np 4 --cache-ram 2048 -ctk q4_0 -ctv q4_0 -fa on -ngl 999 --jinja --alias qwen3.8-9b-distill; no --reasoning-format on the line | no --profile flag: each suite's default (extraction for most; agentic-fast for toolscale, tools, chainwalk; study-nothink for epistemics) | the gauntlet Alan asked for |
qwen3.8-9b-distill-recheck | 2026-09-17 ~17:35Z start; screening finished 17:40:52Z (run-log 18:40:52 BST) | llama.cpp/build/bin/llama-server, manifest b10893-c32d1dabe | -c 8192 -np 1 -ngl 99 --jinja --no-warmup, KV default (f16) | --profile extraction | one of six spot-checks of whether the board reproduces |
Gauntlet driver: a one-off script looped over toolcall toolscale tools nullschema worldview synthesis screening grounding epistemics chainwalk contract enrichment tagger tagger_nullable, each as uv run --quiet --python 3.12 python -m harness.runner --suite $s --model-id qwen3.8-9b-distill --label qwen3.8-9b-distill. enable_thinking: false is sent by the harness profile; the first test with 16 max_tokens returned empty content only because reasoning used the budget.
Other servings with no run-log label: the 3-run speed probes of 2026-09-14, the quiet-box probe (30.5 tok/s) and the overnight probe with -c 16384 -np 1 --cache-ram 1024 -ctk q4_0 -ctv q4_0 (30.8 tok/s decode, 397.9 prefill, 9.3 GB).
Seats held: one, from 2026-09-14 14:30Z, replacing agents-a1. The server was stopped no later than the overnight speed script of 23:02Z, which stops that server once gauntlets clear; the exact stop time is not on record. It was never a router preset or fleet seat.
Decisions
- Downloaded on Alan's request; Q8_0 chosen by the assistant.
- Not seated: the seat plan of 2026-09-14 17:21Z did not become a preset; the file was classed REFERENCE on 2026-09-16.
- Kept through the 2026-10-11 sweeps;
models/qwen3.8-9b-distill/is present on 2026-10-23.
What went wrong
- Void warning on toolcall. The harness printed a warning to treat the run as void until the responses were inspected, because the arms of the suite scored identically. The assistant inspected: 60 of 60 answered directly, none empty,
finish_reasonstop; a direct call with a tool returnedtool_calls; a replay of case 0 reproduced it. The result stands; the run log records it as not void. - Contended speed readings. 14.4 / 14.4 / 11.9 tok/s were taken with the gauntlet running; 30.5 is the clean figure.
- A cross-machine claim was withdrawn the same day. On 2026-09-14 16:37Z the assistant noted that the 26.2 tok/s M5 Pro figure it had compared against "came from the MLX card", so its efficiency comparison mixed runtimes. This concerned the Bonsai comparison, not this file's numbers.
- Server settings differ from the 27B baselines. This label ran with
-np 4and q4_0 KV; the 27B labels it is compared with ran-np 1and q8_0 KV (qwen3.8-27b-q4_0,qwen38-27b-np1), and the fleet notes that-np 2against-np 1flipped 7 worldview and 8 nullschema cases on another model. The recheck at-np 1, f16 KV reproduced screening exactly, but no other suite was re-run. Cross-model rows should be read with that caveat. - Profile mix. toolscale, tools and chainwalk ran under
agentic-fastand epistemics understudy-nothinkbecause no profile was passed; the 27B comparison rows for toolscale ran underextraction. - Stale build stamps: none. Run-log
runtime_versionb10893 (c32d1dabe) agrees with the manifest and with--version. Partial runs: none (row counts in the family doc). Missing flags:--jinjapresent; no--reasoning-format; thinking disabled per request. - Failed Q4_0 build left a 10 MB stub (see the stub doc).
What happened to it
Downloaded 2026-09-14 14:17Z to 14:25Z → benched 14:31Z to 15:45Z → speed probes 2026-09-14/15 → screening recheck 2026-09-17 → classed REFERENCE 2026-09-16 → on disk on 2026-10-23, 9.79 GB. Not deleted by any sweep; no RESTORE entry is needed. Re-downloadable from empero-ai/Qwen3.8-9B-Distill-GGUF at revision 760121cd.