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

Q8_0

A near-full-precision copy, and the first we tested. Run the same way, it scored the same as the original on every test. Quantised here from IFM's BF16.

Size: 5.4 GB
Status: Deleted 12 Sep 2026
Scored runs: 13
Smallest Mac: 16 GB estimate, 8K context

Results13 scored · 95% intervals

Suite	Score	95% interval	Passed	Runtime · profile	Recorded
chainwalkk2-horizon-3.7b	50.0%	9.5–90.5	Passed 1/2	llama.cppextraction	2026-09-09
nullschemak2-3.7b-q8-precision	43.1%	32.3–54.6	Passed 31/72	llama.cppextraction	2026-09-10
nullschemak2-horizon-3.7b	43.1%	32.3–54.6	Passed 31/72	llama.cppextraction	2026-09-09
screeningk2-horizon-3.7b	100.0%	94.0–100.0	Passed 60/60	llama.cppextraction	2026-09-09
synthesisk2-3.7b	58.3%	45.7–69.9	Passed 35/60	llama.cppextraction	2026-09-09
synthesisk2-3.7b-q8-precision	56.7%	44.1–68.4	Passed 34/60	llama.cppextraction	2026-09-10
toolcallk2-3.7b-q8-precision	100.0%	94.0–100.0	Passed 60/60	llama.cppextraction	2026-09-10
toolcallk2-horizon-3.7b	100.0%	94.0–100.0	Passed 60/60	llama.cppextraction	2026-09-09
toolsk2-horizon-3.7b	86.7%	62.1–96.3	Passed 13/15	llama.cppextraction	2026-09-09
toolscalek2-3.7b-q8-precision	100.0%	91.2–100.0	Passed 40/40	llama.cppextraction	2026-09-10
toolscalek2-horizon-3.7b	100.0%	91.2–100.0	Passed 40/40	llama.cppextraction	2026-09-09
worldviewk2-3.7b	50.0%	35.2–64.8	Passed 20/40	llama.cppextraction	2026-09-09
worldviewk2-3.7b-q8-precision	50.0%	35.2–64.8	Passed 20/40	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-Q8_0.gguf`
Format	Q8_0, 5,386,404,224 B (HF size of the same-type quant; 5.0 GiB / 5.39 GB local)
Where from	made here 2026-09-09 from the BF16, fork `llama-quantize`
Status	deleted 2026-09-12 08:04Z
Hugging Face	not downloaded; same-type file published as `K2-Horizon-4B-Q8_0.gguf`

How it was made

Input: models/k2-horizon-3.7b/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-Q8_0.gguf Q8_0

Tool: k2horizon fork @ 35999d1 (build 20). No imatrix.
Date: 2026-09-09, between ~16:02Z (BF16 metadata) and 16:03Z (file listed).
Who asked: the assistant's script, under Alan's push for the 4B. sha256 not on record.

How it was served

label	dates (UTC)	binary and version	ctx / -np / KV / -fa / `--jinja` / `--reasoning-format`	harness profile	why this label exists
`k2-horizon-3.7b`	2026-09-09 16:04Z–16:16Z	k2horizon fork b20 `35999d1`	`-c 32768` / -np default / f16 KV (default) / -fa default / `--jinja` yes / none; `-ngl 999`	extraction	first gauntlet. The script passed `--label k2-horizon-3.7b-q8`, but the runner did not yet honour `--label`, so results were filed under the model id (label fix verified 16:57Z)
`k2-3.7b`	2026-09-09 16:52Z (synthesis/1), 17:40Z (synthesis/2), 18:28Z (worldview/1, superseded), 18:42Z (worldview/2)	k2horizon fork b20	`-c 8192 -np 4` (2,048 per slot) / f16 KV / -fa default / `--jinja` yes / none; `-ngl 999`	extraction	the co-resident fleet seat, used for the cross-fleet synthesis and worldview runs
`k2-3.7b-q8-precision`	2026-09-10 11:16Z–11:24Z	k2horizon fork b20	`-c 32768 -np 1` / `-ctk q8_0 -ctv q8_0` / `-fa on` / `--jinja` yes / `deepseek`; `-ngl 999`	extraction	middle rung of the precision ladder

Unlabelled measurements on this file: the n=32 concurrency test (-c 65536 -np 32 --jinja).

Seats held (reconstructed from launch commands; no /props capture names the file):

Fleet seat 2026-09-09 16:30Z → 2026-09-11 13:59Z (-c 8192 -np 4; /props on 2026-09-11 09:22Z: build b20-35999d1 n_ctx 2048 n_slots 4).
opencode small_model (compaction) until 2026-09-09 20:25Z.
Hermes aux title_generation, profile_describer from 2026-09-10 16:52Z.

Decisions

Made as the near-lossless serving quant alongside the BF16 download.
Seated for the co-resident fleet.
Replaced by the Q4_0 at the 2026-09-11 15:53Z fleet restart.
Deleted 2026-09-12 08:04Z in Alan's "kill and q8 and q4_k_m safely?" sweep. The assistant warned that the Q8_0 references went with it.
Not published.

What went wrong

Label not applied. --label k2-horizon-3.7b-q8 was ignored by the runner of the time, so the first gauntlet is filed as k2-horizon-3.7b.
Stale build stamps. k2-horizon-3.7b rows say build 10618 (eb25b7263); k2-3.7b worldview/2 (2026-09-09) and all k2-3.7b-q8-precision rows say build 10893, a build that did not exist until 2026-09-10. All were served by the fork.
2,048-token slots on the seat (flagged). No run was truncated: the largest prompt+completion in the synthesis/worldview files is 434 tokens (result files, 2026-10-23).
Thinking budget. It "spent 331 tokens answering 'what is the capital of Scotland'" with reasoning_effort=high mandated; small max_tokens probes return empty content.

What happened to it

Made 2026-09-09 → benched 2026-09-09/10 → seat 2026-09-09 → 2026-09-11 → deleted 2026-09-12 08:04Z, no RESTORE entry → recreatable with the command above from the BF16, or downloadable as IFM's K2-Horizon-4B-Q8_0.gguf (a different artefact unless shown byte-identical).