K2-Horizon-7B · file← K2-Horizon-7B

Q8_0

Made here, and the one most of the testing ran on, including its three hours in service. Quantised here from IFM's BF16.

Size: 9.6 GB
Status: Deleted 12 Sep 2026
Scored runs: 12
Smallest Mac: 18 GB estimate, 8K context

Results12 scored · 95% intervals

Suite	Score	95% interval	Passed	Runtime · profile	Recorded
chainwalkk2-horizon-7b	50.0%	9.5–90.5	Passed 1/2	llama.cppextraction	2026-09-09
contractk2-horizon-7b	22.9%	13.3–36.5	Passed 11/48	llama.cppextraction	2026-09-09
groundingk2-horizon-7b	55.2%	46.5–63.6	Passed 69/125	llama.cppextraction	2026-09-09
nullschemak2-7b-q8-precision	36.1%	26.0–47.6	Passed 26/72	llama.cppextraction	2026-09-10
nullschemak2-horizon-7b	34.7%	24.8–46.2	Passed 25/72	llama.cppextraction	2026-09-09
screeningk2-horizon-7b	90.0%	79.9–95.3	Passed 54/60	llama.cppextraction	2026-09-09
synthesisk2-7b	75.0%	62.8–84.2	Passed 45/60	llama.cppextraction	2026-09-09
toolcallk2-horizon-7b	95.0%	86.3–98.3	Passed 57/60	llama.cppextraction	2026-09-09
toolsk2-horizon-7b	100.0%	79.6–100.0	Passed 15/15	llama.cppextraction	2026-09-09
toolscalek2-horizon-7b	100.0%	91.2–100.0	Passed 40/40	llama.cppextraction	2026-09-09
worldviewk2-7b	30.0%	18.1–45.4	Passed 12/40	llama.cppextraction	2026-09-09
worldviewk2-7b-q8-precision	30.0%	18.1–45.4	Passed 12/40	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-Q8_0.gguf`
Format	Q8_0, 8.92 GiB (≈9.57 GB)
Where from	made here 2026-09-09 from the BF16, fork `llama-quantize`
Status	deleted 2026-09-12 08:04Z
Hugging Face	not downloaded (HF had BF16 only on 2026-09-09)

How it was made

Input: models/k2-horizon-7b/K2-Horizon-7B-BF16.gguf (see K2-Horizon-7B BF16 (IFM's published GGUF)).
Command (the run that completed is the relaunch logged to a scratch file):

llama.cpp-k2horizon/build/bin/llama-quantize \
models/k2-horizon-7b/K2-Horizon-7B-BF16.gguf \
models/k2-horizon-7b/K2-Horizon-7B-Q8_0.gguf Q8_0

No imatrix and no tensor-type overrides.

Tool: MBZUAI-IFM fork model/K2Horizon @ 35999d1 (build 20), built 2026-09-09 14:54Z.
Date: 2026-09-09, complete by 15:11:46Z.
Who asked: Alan, "we could quantize q8 and run it, see what it can do against the gauntlet", then "the 7b model looks more interesting".

How it was served

All on the fork's llama-server (llama.cpp-k2horizon/build/bin/llama-server, b20 / 35999d1).

label	dates (UTC)	ctx / -np / KV / -fa / --jinja / --reasoning-format	profile	why this label exists
`k2-horizon-7b`	2026-09-09 15:12Z–15:38Z	gauntlet: `--jinja -c 32768 -ngl 999` (no `-np`, f16 KV, no `-fa`, no `--reasoning-format`); walk: same with `-c 131072`	extraction (thinking off, max_tokens 2048)	first gauntlet: toolcall, contract, screening, nullschema, grounding, then toolscale, chainwalk, tools. The scripts passed `--label k2-horizon-7b-q8`, but files and run-log rows were filed under the model id `k2-horizon-7b` (see Incidents)
`k2-7b`	2026-09-09 16:50Z–18:43Z	`--jinja -c 8192 -np 2 -ngl 999 --alias k2-7b` → 4096 per slot; f16 KV	extraction	the co-resident fleet seat; synthesis/1, synthesis/2 and worldview/2 ran against it from the fleet-wide scripts
`k2-7b-q8-precision`	2026-09-10 11:43Z–11:48Z	`--jinja --reasoning-format deepseek -c 32768 -np 1 -fa on -ctk q8_0 -ctv q8_0 -ngl 999`, alias k2-7b-q8	extraction (temp 0, seed 42)	middle rung of the precision ladder

Seat held: 2026-09-09 16:30Z → 19:43Z, in opencode and pi, "K2-Horizon-7B (8.9GB) - 200-step chains". Stood down for the 36B. The 4096-per-slot window did not bind on these suites: the longest prompt in its synthesis/worldview files is 288 tokens, and every case finished stop (results files, read 2026-10-23).

Decisions

Made as the build for Alan's gauntlet request.
Seated in the co-resident fleet after Alan's "it would need paired with another model". Stood down 19:43Z as "the weakest of the four on most measures".
Deleted 2026-09-12 08:04Z as one of the "Q8_0 reference builds", on Alan's "We can do some more cleanup too, kill and q8 and q4_k_m safely?". Its Q4_K_M sibling was kept.
Not published.

What went wrong

--label ignored. Both gauntlet scripts passed --label k2-horizon-7b-q8, and the synthesis/1 script passed --label synthesis-k2-7b. The files were written under the model id instead: "no matches found: toolcall__k2-horizon-7b-q8__*.jsonl". A lab note records that runner.py "gained --label" later that day (reconstructed: the flag was not yet honoured). Consequence: the run-log label k2-horizon-7b names no quant.
Partial worldview/1 run. worldview__k2-7b__extraction.jsonl.superseded holds 5 of 40 rows (mtime 19:29 local = 18:29Z) of worldview/1, the first version, which embedded internal infrastructure notes. Alan interrupted at 18:29:00Z ("you don't need 140 docs for this"... "some of old is private"). The suite was rebuilt as worldview/2 without the infrastructure excerpts and re-run (reconstructed: the timestamps match). The partial file is not in the run log.
Gauntlet server flags: no -np, so the slot count was the fork's default (not on record); f16 KV; no --reasoning-format. The label's intent (temperature 0, thinking off) is consistent with that, and no reasoning content was separated (reasoning_chars 0 throughout).
Run-log build stamps: the k2-horizon-7b and synthesis entries say upstream build 10618 (eb25b7263). The k2-7b worldview entry (2026-09-09) says build 10893 (c32d1dabe), a build that did not exist on this box until 2026-09-10 12:17Z. All of these runs were served by the fork b20 (35999d1); runtime_version is right and llama_build is wrong.

What happened to it

Quantised here 2026-09-09 → benched 2026-09-09 (8 + 3 runs) and 2026-09-10 (2) → seated 2026-09-09 16:30Z–19:43Z → deleted 2026-09-12 08:04Z. No RESTORE-deleted-*.sh entry. Re-create: download K2-Horizon-7B-BF16.gguf from IFM/K2-Horizon-7B-GGUF and run the command above (fork, or upstream once it is verified on this arch), or take IFM's later-published Q8_0, which is a different artefact unless verified identical. The run-log rows survive.