Q4_0

Unsloth's 4-bit build, run beside NeoHorse as the control. Seven tests, then deleted. Unsloth GGUF.

Size: 2.6 GB
Status: Deleted 11 Oct 2026
Scored runs: 7
Smallest Mac: 8 GB estimate, 8K context

Results7 scored · 95% intervals

Suite	Score	95% interval	Passed	Runtime · profile	Recorded
epistemicsqwen3.5-4b-base	85.4%	72.8–92.8	Passed 41/48	llama.cppstudy-nothink	2026-09-13
groundingqwen3.5-4b-base	74.4%	66.1–81.2	Passed 93/125	llama.cppextraction	2026-09-13
nullschemaqwen3.5-4b-base	33.3%	23.5–44.8	Passed 24/72	llama.cppextraction	2026-09-13
tagger_nullableqwen3.5-4b-base	93.8%	89.9–96.2	Passed 225/240	llama.cppextraction	2026-09-13
toolcallqwen3.5-4b-base	36.7%	25.6–49.3	Passed 22/60	llama.cppextraction	2026-09-13
toolscaleqwen3.5-4b-base	100.0%	91.2–100.0	Passed 40/40	llama.cppextraction	2026-09-13
worldviewqwen3.5-4b-base	40.0%	26.3–55.4	Passed 16/40	llama.cppextraction	2026-09-13

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.

The recordHow it was made · how it was served · what went wrong · what happened to it

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	`models/qwen3.5-4b-base/Qwen3.5-4B-Q4_0.gguf`
Format	GGUF Q4_0, 2,583,221,408 B
Where from	downloaded 2026-09-13 ~06:52Z from `unsloth/Qwen3.5-4B-GGUF` with a 6-range `pget.sh` that checks the byte count against `content-length`
Status	deleted 2026-10-11 (Tier B sweep); comment-only RESTORE entry
Hugging Face	https://huggingface.co/unsloth/Qwen3.5-4B-GGUF

Where it came from

Repo / file: unsloth/Qwen3.5-4B-GGUF / Qwen3.5-4B-Q4_0.gguf, from resolve/main (revision not pinned).
Command:

cd models/qwen3.5-4b-base
$SP/pget.sh "https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/main/Qwen3.5-4B-Q4_0.gguf" Qwen3.5-4B-Q4_0.gguf 6

pget.sh fetches six byte ranges and refuses the file unless its size equals the server's content-length. The first attempt with hf download failed because hf was not on PATH.

Landed: 06:55Z, "2.41 GB" (GiB) on disk.
Size: 2,583,221,408 B (source: hf; the download's own size check agrees to the GiB shown). sha256 not on record.
Who asked / why: the assistant, as NeoHorse's control arm, inside Alan's request to benchmark NeoHorse.
Directory name qwen3.5-4b-base/ and label qwen3.5-4b-base mean "NeoHorse's base". The weights are the post-trained Qwen/Qwen3.5-4B, not Qwen/Qwen3.5-4B-Base (inference from the repo name unsloth/Qwen3.5-4B-GGUF; unsloth's card was not read).

How it was served

label	dates (UTC)	binary and version	ctx / -np / KV / -fa / `--jinja` / `--reasoning-format`	harness profile	why this label exists
`qwen3.5-4b-base`	2026-09-13 ~07:50–08:10	`.local/bin/llama-server` → `llama.cpp`, b10893 (`c32d1dabe`), from /props	16384 / 1 / f16 (default) / on / on / not set	extraction (6 suites), study-nothink (epistemics)	the NeoHorse control arm; profiles chosen per suite to match the board's dominant profile "so the results are comparable rather than a new island"

Command:

$LS -m models/qwen3.5-4b-base/Qwen3.5-4B-Q4_0.gguf --jinja \
    -c 16384 -np 1 -fa on -ngl 999 --alias qwen3.5-4b-base
python harness/runner.py --suite <s> --model <server> --model-id qwen3.5-4b-base \
    --label qwen3.5-4b-base --profile extraction        # study-nothink for epistemics

Run sequentially after NeoHorse on the same port; the box held two other seats (gemma4-12b and ornith-1.5-9b) during the run. Correctness suites were judged host-independent.

Seats held: none. Never in fleet-models.ini, fleetview or launchd. Listed in seat-planning tables only (2026-09-14 seat-cost table row: qwen3.5-4b-base, Q4_0, dense, 2.6G weights, 128.0K context, 2.15G KV, 4.7G seat, 2026-09-16 "fastest format present").

Decisions

Downloaded as the control arm.
Not seated, not published. Its role was the comparison, and the result was written up only in-session.
Deleted 2026-10-11 in the Tier B sweep, "benched-once, abandoned" (a lab note).

What went wrong

Run started ~30 minutes late because the assistant waited for a go-ahead it had already been given. No cost beyond time.
Standing checks: no partial runs; run-log build b10893 matches the binary the evidence shows; --jinja present; reasoning format not set, consistent with the extraction/study-nothink intent (thinking off).

What happened to it

Downloaded 2026-09-13 → benched 2026-09-13 (7 suites, one run each) → never seated → deleted 2026-10-11 (Tier B; RESTORE-deleted-20261011-tiersAB.sh has a comment line but no download command) → re-downloadable from unsloth/Qwen3.5-4B-GGUF (Qwen3.5-4B-Q4_0.gguf; current revision may differ).