Qwen3.8-27B · file← Qwen3.8-27B
UD-Q4_K_XL
The first file; our first tests ran on it. Deleted when we moved to MLX. Unsloth GGUF + mmproj.
- Size
- 17.9 GB
- Status
- Deleted 9 Sep 2026
- Scored runs
- 5
- Smallest Mac
- 32 GB estimate, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| epistemicsqwen3.8-27b | 100.0% | 92.6–100.0 | Passed 48/48 | llama.cppstudy-nothink | 2026-08-24 |
| groundingqwen3.8-27b | 80.0% | 72.1–86.1 | Passed 100/125 | llama.cppextraction | 2026-08-24 |
| groundingqwen3.8-27b-noreason | 80.0% | 72.1–86.1 | Passed 100/125 | llama.cppextraction | 2026-09-07 |
| toolcallqwen3.8-27bvoid | 100.0% | 94.0–100.0 | Passed 60/60 | omlxextraction | 2026-09-09 |
| toolsqwen3.8-27b | 100.0% | 79.6–100.0 | Passed 15/15 | 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-Q4_K_XL.gguf (+ mmproj-F16.gguf) |
| Format | Q4_K_XL, about 17.9 GB (bytes not recorded) |
| Where from | downloaded on or before 2026-08-14 |
| Status | router seat, then deleted 2026-09-09 (RESTORE-deleted-20260909.sh) |
| Hugging Face | https://huggingface.co/unsloth/Qwen3.8-27B-GGUF |
Where it came from
- Repo / file:
unsloth/Qwen3.8-27B-GGUF/Qwen3.8-27B-UD-Q4_K_XL.gguf, kept inmodels/qwen3.8-27b/withmmproj-F16.ggufbeside it (fleet-models.ini preset; RESTORE-deleted-20260909.sh). - Date downloaded: not on record (would have been in a session before 2026-08-23, or the shell history of the Mac Studio). It was already installed on 2026-08-14: the router preset block is headed 'NEW 2026-08-14' (
fleet-models.ini, commented history), and our agents' test of 2026-08-14 used it through the router. The earliest session on record is from 2026-08-23; a probe for a download script in this directory that day found none (the cat fell through to another model's script), so the method used is not on record. - Size: about 17.92 GB local (assistant's table), 5.31 bits per weight (fleet-models.ini comment). HF's current file is 17,559,178,144 B, which differs, so the local copy was an older unsloth upload. Exact local bytes and sha256 are not on record.
- HF revision: not on record.
- Companion:
mmproj-F16.gguf(vision projector). Alan asked what dense Qwen3.8 models existed on 2026-08-23 and was told the 27B is the only dense one, natively multimodal.
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
qwen3.8-27b (tools, epistemics) | 2026-08-24 00:37–00:41 | llama.cpp llama-server, build b10369 (6e62ba538, 0.2.0-dev) | -ngl 99 -c 32768 -np 2 -cb --jinja; no -fa, no KV type, no --reasoning-format, no spec-type (launch time (reconstructed) from the run-log stamps) | agentic-fast (tools), study-nothink (epistemics) | the 'for completeness' screen on the suites that had just been shown to be saturated, queued so it could run overnight (Alan: "do the 3.8 at the end when I say so it can run overnight") |
qwen3.8-27b (grounding) | 2026-08-24 13:13 (finish) | same binary | -ngl 99 -c 32768 -np 4 -cb --jinja, nothing else (grounding sweep script) | extraction | first row of the grounding/1 sweep across six candidates; waited behind the Toolery sweep for the GPU |
qwen3.8-27b-noreason (grounding) | 2026-09-07 20:46 | b10618 (eb25b7263, rebuilt 2026-08-25) | --ctx-size 16384 --n-gpu-layers 99 --jinja | extraction | control for the MLX result: same file, no reasoning kwargs, no MTP, to see whether 'reasoning effort' explained the 12-point MLX lead |
qwen3.8-27b (toolcall, 2026-09-09 10:56) | — | oMLX 0.6.4 | — | extraction | not this file. That row served qwen3.8-27b-mtp-medium/, an MLX build; see Qwen3.8-27B MLX 4-bit with MTP sidecar (rapid-mlx curated build; served as qwen3.8-27b-mtp-medium). It is voided in the run log. |
Seats held: the router preset [qwen3.8-27b] (spec-type draft-mtp, ctx 131072 cut to 65536 on 2026-08-19 after a dsh session died at the ceiling; fleet-models.ini comments), used by Hermes, until it was retired on 2026-09-09 (fleet-models.ini 'RETIRED 2026-09-09'). The router preset is the only place MTP speculation and the reasoning-effort kwargs were configured for this file; none of the run-log rows above used them (the launch lines contain no spec-type and no chat-template-kwargs).
Decisions
- Why it was benched at all: completeness. Alan on 2026-08-23: "3.8 is good I don't doubt it, too slow for what I need to do and qwen models excel at thinking not doing lol"; he had called it funny that "everyone is talking about how amazing qwen3.8 is and I'm going back a .5 - 1 gen in models".
- Why it was kept: it was the router's Qwen3.8 build and the baseline for the MLX comparison. On 2026-09-04 the assistant noted it was the only MLX model with a GGUF counterpart that had suite baselines, which is what made it the test subject on 2026-09-07.
- Why it was deleted (2026-09-09): Alan, 14:29Z: "the guff models can be removed for qwen3 next and 3.8 as we have mlx now". The assistant checked the vision path first and flagged that this left Muse-Glimmer as the only other vision-capable model.
- Published: the model write-up
qwen3.8-27bon the site; this file's rows are theqwen3.8-27band-noreasonlabels in the run log.
What went wrong
- Mixed label (published-data mismatch).
qwen3.8-27bis this GGUF for tools/epistemics/grounding and an MLX build for the toolcall row. Anything keyed on the label alone mixes two weights files. - First
-noreasonattempt was killed at 19 rows. The foreground launch hit the tool's 2-minute limit (exit 143); the assistant wrote that the group died 'before the suite ran', but a 19-row grounding file exists, nowgrounding__qwen3.8-27b-noreason__extraction.jsonl.superseded(last row 21:39 BST). The re-run in the background produced the 125-row file. Cost: about 7 minutes; no wrong number published. - Baseline mislabelled in the narration. On 2026-09-07 the assistant tabulated the 80.0% grounding baseline as 'GGUF UD-Q4_K_XL + xhigh + MTP'. The 2026-08-24 sweep script launched a plain llama-server with no spec-type and no chat-template-kwargs, so the baseline was the same configuration as the control, and both scoring 80.0 is unsurprising. Reasoning effort was never the variable for that row.
- Toolery contention. The 1-bit run and the Toolery sweep fought for the GPU on 2026-08-24; Toolery numbers for this model are timeouts, not capability (above).
- Missing flags (standing check): no
-fa, no KV type and no--reasoning-formaton any launch line;--jinjapresent. No stale build stamps: the run log's b10369 and b10618 match the evidence.
What happened to it
Downloaded on or before 2026-08-14 → router seat (Hermes) → benched 2026-08-24 → controlled 2026-09-07 → deleted 2026-09-09 about 14:28Z with a RESTORE-deleted-20260909.sh entry (18 GB; the script text says to re-download with an include pattern for UD-Q4_K_XL plus mmproj-F16 into qwen3.8-27b, then uncomment the preset). Re-downloadable from unsloth/Qwen3.8-27B-GGUF; the current file is not byte-identical to the one benched.