K2-Horizon-3.7B · file← K2-Horizon-3.7B

Q4_K_M

The usual 4-bit format, to see what smaller costs. Deleted once we had made the Q4_0 from it. Quantised here from IFM's BF16.

Size: 3.2 GB
Status: Deleted 12 Sep 2026
Scored runs: 5
Smallest Mac: 8 GB estimate, 8K context

Results5 scored · 95% intervals

Suite	Score	95% interval	Passed	Runtime · profile	Recorded
nullschemak2-3.7b-q4-precision	38.9%	28.5–50.4	Passed 28/72	llama.cppextraction	2026-09-10
synthesisk2-3.7b-q4-precision	50.0%	37.7–62.3	Passed 30/60	llama.cppextraction	2026-09-10
toolcallk2-3.7b-q4-precision	100.0%	94.0–100.0	Passed 60/60	llama.cppextraction	2026-09-10
toolscalek2-3.7b-q4-precision	100.0%	91.2–100.0	Passed 40/40	llama.cppextraction	2026-09-10
worldviewk2-3.7b-q4-precision	52.5%	37.5–67.1	Passed 21/40	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	`K2-Horizon-3.7B-Q4_K_M.gguf`
Format	Q4_K_M, 3,156,598,144 B (HF size; 2.9 GiB / 3.16 GB local)
Where from	made here 2026-09-09 from the BF16
Status	deleted 2026-09-12 08:04Z
Hugging Face	not downloaded; same-type file published as `K2-Horizon-4B-Q4_K_M.gguf`

How it was made

Input: K2-Horizon-4B-BF16.gguf (see K2-Horizon-3.7B BF16 (K2-Horizon-4B-BF16.gguf, downloaded from IFM)).
Command:

llama.cpp-k2horizon/build/bin/llama-quantize K2-Horizon-4B-BF16.gguf \
models/k2-horizon-3.7b/K2-Horizon-3.7B-Q4_K_M.gguf Q4_K_M

Tool: k2horizon fork @ 35999d1; no imatrix. Date: 2026-09-09 ~16:03Z.
GGUF header read from this file: dense, 36 blocks, embedding 2560, 32/8 heads, FFN 10240.

How it was served

label	dates (UTC)	binary and version	ctx / -np / KV / -fa / `--jinja` / `--reasoning-format`	harness profile	why this label exists
`k2-3.7b-q4-precision`	2026-09-10 11:06Z–11:15Z	k2horizon fork b20 `35999d1`	`-c 32768 -np 1` / `-ctk q8_0 -ctv q8_0` / `-fa on` / `--jinja` yes / `deepseek`; `-ngl 999`	extraction	bottom rung of the precision ladder

Unlabelled uses: draft model for K2-36B speculative decoding — 2026-09-09 23:37Z+ (--model-draft, failed on a renamed flag), and 2026-09-11 14:25Z (-md... --spec-type draft-simple --spec-draft-n-max 4 -c 32768 -np 1 -fa on -ctk q8_0 -ctv q8_0 --reasoning-format deepseek). Never seated.

Decisions

Made with the Q8_0 as the board's standard 4-bit format.
Kept on 2026-09-11 15:14Z as a "drafter/MTP" file under Alan's "kill what you can disk wise, keep any q4_0 and drafters", after it had already been requantised to Q4_0.
Deleted 2026-09-12 08:04Z as a Q4_K_M "with a Q4_0 sibling".

What went wrong

Stale build stamp: run-log llama_build build 10893 on all five rows; the fork served them.
Contended speed number never re-measured.
First speculation attempt (2026-09-09) failed because --draft-max had been renamed.

What happened to it

Made 2026-09-09 → benched 2026-09-10 → draft model 2026-09-09/11 → source of the Q4_0 2026-09-11 → deleted 2026-09-12 08:04Z, no RESTORE entry → recreatable from the BF16 with the command above; IFM publishes a same-type file (K2-Horizon-4B-Q4_K_M.gguf), not shown identical.