Spark-X2.5-4B · file← Spark-X2.5-4B

Q4_K_M

XHToken's own 4-bit file. Tested once on 10 September, never given a job, deleted on 11 October. XHToken's GGUF, downloaded 2026-09-10, deleted 2026-10-11.

Size: 2.6 GB
Status: Deleted 11 Oct 2026
Scored runs: 5
Smallest Mac: 8 GB estimate, 8K context

Results5 scored · 95% intervals

Suite	Score	95% interval	Passed	Runtime · profile	Recorded
nullschemaspark-x25-4b	48.6%	37.4–59.9	Passed 35/72	llama.cppextraction	2026-09-10
synthesisspark-x25-4b	28.3%	18.5–40.8	Passed 17/60	llama.cppextraction	2026-09-10
toolcallspark-x25-4b	0.0%	-0.0–6.0	Passed 0/60	llama.cppextraction	2026-09-10
toolscalespark-x25-4b	95.0%	83.5–98.6	Passed 38/40	llama.cppextraction	2026-09-10
worldviewspark-x25-4b	20.0%	10.5–34.8	Passed 8/40	llama.cppextraction	2026-09-10

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	`Spark-X2.5-4B-Q4_K_M.gguf`
Format	Q4_K_M GGUF, 2,600,224,352 B (HF listing; 2.60 GB on disk 2026-09-12)
Where from	downloaded 2026-09-10 from `XHToken/Spark-X2.5-4B-GGUF` (6 parallel curl range requests)
Status	deleted 2026-10-11 (tier B sweep)
Hugging Face	https://huggingface.co/XHToken/Spark-X2.5-4B-GGUF

Where it came from

Repo / file: XHToken/Spark-X2.5-4B-GGUF / Spark-X2.5-4B-Q4_K_M.gguf, the vendor's own GGUF repo. A bartowski lookup returned HTTP 401 first.
Command, into models/spark-x25-4b/:

URL=https://huggingface.co/XHToken/Spark-X2.5-4B-GGUF/resolve/main/Spark-X2.5-4B-Q4_K_M.gguf
SIZE=$(curl -sIL "$URL" | awk -F': ' '/^[Cc]ontent-[Ll]ength/{print $2}' | tail -1 | tr -d '\r')
# 6 parallel range requests (curl -sL -r S-E "$URL" -o part.$i), then
cat part.* > Spark-X2.5-4B-Q4_K_M.gguf && rm -f part.*

Revision: not on record (resolve/main, unpinned).
sha256: not on record. The sha … values in the speed log hash the responses, not the file.
Date: started 2026-09-10 07:46Z; complete by 08:48Z (the bench log reads the file at 2.42 GB). Exact finish time not on record.
Who asked: Alan, "Spark x2.5 another model to look at" / "Sure".
GGUF header: general.architecture = spark2_5, vocab 131,072. A later provenance survey said of this directory "No fingerprints and no evidence either way". The download command above settles it.

How it was served

label	dates (UTC)	binary and version	ctx / -np / KV / -fa / `--jinja` / `--reasoning-format`	harness profile	why this label exists
— (attempt 1, no results)	2026-09-10 ~08:47Z	upstream `llama.cpp` b10618 (eb25b7263)	`-c 131072 -np 1 -fa on -ctk q8_0 -ctv q8_0 -ngl 999 --jinja --reasoning-format deepseek` (inherited from a one-off script)	extraction (planned)	first queue position on the same-day board; failed to load
`spark-x25-4b`	2026-09-10 12:34Z–12:40Z	upstream `llama.cpp` b10893 (c32d1dabe), built that day	`-c 32768 -np 1 -fa on -ctk q8_0 -ctv q8_0 -ngl 999 --jinja --reasoning-format deepseek`, `--alias spark-x25-4b`	extraction (temperature 0, thinking off, 2048 max_tokens, seed 42; manifest)	the board row, once b10893 knew `spark2_5`

The -c 32768 follows the "mid-range models run at -c 32768 rather than 131072" decision, defended on 32k/f16 vs 128k/q8 giving identical worldview verdicts.
Run-log runtime llama.cpp b10893 (c32d1dabe), runtime_source "evidence"; llama_build 0.4.0-dev (build 10893, commit c32d1dabe) agrees.
Seats held: none. It never had a router preset or a fleetview entry.

Decisions

Downloaded as the only Spark build; the vendor's Q4_K_M was the smallest file offered.
Benched once; never seated. Verdict: "a dispatcher, not a router".
Kept on 2026-09-12 as the sole build of its model and on 2026-09-16.
Deleted 2026-10-11 in tier B ("benched-once, abandoned") on Alan's "nuke tiers a & b".
Not published to HF (third-party file).

What went wrong

Unknown architecture on b10618. llama_model_load: error loading model: unknown model architecture: 'spark2_5'. Support had landed upstream as b10828 on 2026-09-06, but the checkout was b10618 from 25 August. Cost: the model missed the morning board and needed a rebuild. Fixed by the b10893 upgrade.
Wrong failure message. The sed-generated script printed "needs the PrismML fork" (Bonsai's fallback text) for an unrelated failure. It also carried Bonsai's label vocabulary (spark-x25-4b-ternary), noticed when the chain reached the GSQ-RCO script. The Spark run itself used a hand-written script with the label spark-x25-4b.
The first launch was blocked by an expired login before the script was written; re-created and launched at 12:34Z.
A zero that was checked. toolcall 0/60 was tested for a parser artefact before being reported.
A false headline. "Fastest model measured" went onto the site card on 2026-09-10 and was corrected on 2026-09-15. An earlier lab note still says "2.4 GB is the fastest model measured".
Standing checks:
partial runs: none;
stale build stamp: none;
missing flags: none (--jinja and --reasoning-format deepseek on the line; profile thinking off by design).

What happened to it

Downloaded 2026-09-10 → failed to load (b10618) 2026-09-10 08:48Z → benched (b10893) 2026-09-10 12:34Z–12:40Z → never seated → kept 2026-09-12 and 2026-09-16 → deleted 2026-10-11 (RESTORE-deleted-20261011-tiersAB.sh has a comment line for it but no re-download command). The run-log results survive.

Re-downloaded 2026-10-24 from revision 9826e0be… as a separate variant doc, Spark-X2.5-4B Q4_K_M. Identity with this file can't be proven: same byte size, no recorded sha256, and the repo was modified on 2026-09-20.