Q4_K_M

A smaller copy made here, scored once in the precision test. The last to be deleted. Quantised here from IFM's BF16.

Size: 5.6 GB
Status: Deleted 11 Oct 2026
Scored runs: 2
Smallest Mac: 16 GB estimate, 8K context

Results2 scored · 95% intervals

Suite	Score	95% interval	Passed	Runtime · profile	Recorded
nullschemak2-7b-q4-precision	37.5%	27.2–49.0	Passed 27/72	llama.cppextraction	2026-09-10
worldviewk2-7b-q4-precision	37.5%	24.2–53.0	Passed 15/40	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-7B-Q4_K_M.gguf`
Format	Q4_K_M, 5,592,217,984 B
Where from	made here 2026-09-09 from the BF16, fork `llama-quantize`
Status	deleted 2026-10-11 (Tier B sweep)
Hugging Face	not downloaded

How it was made

Input: models/k2-horizon-7b/K2-Horizon-7B-BF16.gguf (see K2-Horizon-7B BF16 (IFM's published GGUF)).
Command:

llama.cpp-k2horizon/build/bin/llama-quantize \
models/k2-horizon-7b/K2-Horizon-7B-BF16.gguf \
models/k2-horizon-7b/K2-Horizon-7B-Q4_K_M.gguf Q4_K_M

No imatrix and no tensor-type overrides.

Tool: k2horizon fork @ 35999d1 (build 20).
Date: 2026-09-09, complete by 15:11:46Z.
Size: 5,592,217,984 B, observed on disk 2026-10-07; 5.21 GiB; 5.59 GB decimal. sha256 not on record.
Who asked / why: added by the assistant in the same script, "for a co-resident option" (reconstructed: no request from Alan for this quant is on record).

How it was served

label	dates (UTC)	binary	ctx / -np / KV / -fa / --jinja / --reasoning-format	profile	why this label exists
`k2-7b-q4-precision`	2026-09-10 11:37Z–11:43Z	fork `llama-server` b20 (35999d1)	`--jinja --reasoning-format deepseek -c 32768 -np 1 -fa on -ctk q8_0 -ctv q8_0 -ngl 999`, alias k2-7b-q4	extraction (temp 0, seed 42)	bottom rung of the same-weights precision ladder

Footprint at load: 7.7 GB RSS. Never seated. On 2026-09-11 it was a candidate draft model in the speculative-decoding vocabulary check: same 250,624 vocab as the 36B. It was never run as a draft; the 3.7B was.

Decisions

Kept on 2026-09-12 when the family's Q8_0 went, as "now the ONLY k2-7b build", and again in the 2026-09-16 disk audit ("KEEP... tested, no faster build on disk").
Not made into a Q4_0, unlike the 3.7B and the 36B.
Deleted 2026-10-11 in the Tier B sweep ("benched-once, abandoned"), decision Alan "nuke tiers a & b", executed by one of our agents.
Not published.

What went wrong

The sweep record describes the wrong file's runs. The RESTORE entry and a lab note say "k2-horizon-7b (5.2G): benched 8 rows incl toolcall 95.0, screening 90.0". Those 8 rows (label k2-horizon-7b) were run on the Q8_0, deleted 2026-09-12. The file deleted on 2026-10-11 carried 2 run-log rows (k2-7b-q4-precision).
RESTORE entry has no re-create command. Unlike its neighbours (hf download LiquidAI/LFM2.5-8B-A1B-GGUF), the k2-horizon-7b block is comment-only. The file was made here, so there is nothing to hf download that is known to be identical.
Run-log build stamp: 0.4.0-dev (build 10893, commit c32d1dabe), the upstream binary, against the fork b20 that served it.

What happened to it

Quantised here 2026-09-09 → benched 2026-09-10 (2 suites) → kept 2026-09-12 and 2026-09-16 → deleted 2026-10-11 (Tier B; the restore script has a comment entry, no command). Re-create: BF16 from IFM/K2-Horizon-7B-GGUF plus the command above. IFM's published Q4_K_M matched this file's byte count, but it is not verified identical. The run-log rows survive.