Qwen3.8-27B · file← Qwen3.8-27B
UD-IQ1_S
A 1-bit copy, a third of the size. It fell apart on chained tool calls; deleted. Unsloth 1-bit GGUF.
- Size
- 6.2 GB
- Status
- Deleted 7 Sep 2026
- Scored runs
- 5
- Smallest Mac
- 16 GB estimate, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| chainwalkqwen3.8-iq1s | 0.0% | 0.0–65.8 | Passed 0/2 | llama.cppagentic-fast | 2026-08-24 |
| enrichmentqwen3.8-iq1s79.6% coverage | 60.5% | 45.6–73.6 | Passed 26/43 | llama.cppstudy-nothink | 2026-08-24 |
| epistemicsqwen3.8-iq1s72.9% coverage | 77.1% | 61.0–87.9 | Passed 27/35 | llama.cppstudy-nothink | 2026-08-24 |
| taggerqwen3.8-iq1s44.2% coverage | 31.6% | 21.0–44.5 | Passed 18/57 | llama.cppstudy-nothink | 2026-08-24 |
| toolsqwen3.8-iq1s86.7% coverage | 76.9% | 49.7–91.8 | Passed 10/13 | llama.cppagentic-fast | 2026-08-24 |
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 | Qwen3.8-27B-UD-IQ1_S.gguf |
| Format | IQ1_S, 6,192,222,208 B |
| Where from | downloaded 2026-08-24 |
| Status | benched (partial), deleted 2026-09-07, no RESTORE entry |
| Hugging Face | same repo |
Where it came from
- Repo / file:
unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-IQ1_S.gguf, 6,192,222,208 B (HF size at download time).dushowed 5.77 GiB on 2026-09-07, which is the same number. - Date and tool: 2026-08-24 about 00:45Z, six-stream range download with the
pget.shhelper intomodels/qwen3.8-27b-lowbit/, started in the same command as UD-Q2_K_XL. sha256 and HF revision: not on record. - Who asked: Alan, 00:42Z, "newsflash before look up unsloth dynamic 3.0 update, looks incredibly promising"; then "UD-1bit quants with UD-IQ1_S" and "this is qwen 3.8 27b unsloth quants". The assistant had first advised against 1-bit for his workload, citing unsloth's own warning, then agreed to run it as a memory-matched experiment.
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-iq1s | 2026-08-24, finishes at 02:02 (chainwalk), 02:50 (enrichment), 02:54 (tools), 03:03 (epistemics), 07:02 (tagger, last row) | llama.cpp, b10369 (6e62ba538) | -ngl 99 -c 32768 -np 2 -cb --jinja; no -fa, no KV type | tools and chainwalk agentic-fast; epistemics, enrichment, tagger study-nothink | the 1-bit arm of the memory-matched comparison; the script also listed qwen3.8-q2kxl but that arm never started |
Finish times above are the run log's local stamps minus one hour. Seats held: none.
Decisions
- Stopped by the assistant at 07:18Z because it and the Toolery sweep were fighting over the GPU and the 1-bit question was already answered.
- Alan, 07:18Z: "yesh I think we can drop the 1 bit". The assistant went further and proposed closing the whole low-bit line; no reply from Alan on that is on record.
- Deleted 2026-09-07 17:36Z with the rest of
qwen3.8-27b-lowbit/(15 GB) as 'superseded / measured-and-rejected', after the triage table listed it as 'both tested and poor' and Alan said "Go for it". The 'both tested' statement was wrong for Q2_K_XL, which was never run. - Not published as a model of its own; the rows sit under the
qwen3.8-27bsite write-up's config list.
What went wrong
- Partial run, tagger. Killed at 129 of 240 rows (above). The run-log row says 31.6% on 57 measured, coverage 44% of the rows present.
- GPU contention. The tagger run took 111 s per case partly because Toolery was running at the same time; the 'slow at 1-bit regardless' reading is therefore confounded.
- No stale stamps, no missing-flag problem beyond the usual: no -fa, no KV type.
What happened to it
Downloaded 2026-08-24 00:45Z → benched overnight, stopped 07:18Z → deleted 2026-09-07 17:36Z. No RESTORE-deleted-*.sh entry (grep of the five RESTORE scripts for 'iq1', 'lowbit' and 'qwen3.8-27b' found only the 20260909 entry for the Q4_K_XL). Re-downloadable from unsloth/Qwen3.8-27B-GGUF.