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

Q4_0

The one we kept, our tool caller since September. Made by shrinking the Q4_K_M a second time. Requantised here from the local Q4_K_M.

Size: 3.0 GB
Status: On disk
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-b11552-nocoh	15.3%	8.8–25.3	Passed 11/72	llama.cppextraction	2026-10-11
nullschemak2-3.7b-b11552-noreasonfmt	19.4%	12.0–30.0	Passed 14/72	llama.cppextraction	2026-10-11
nullschemak2-3.7b-fork-control-nocoh	16.7%	9.8–26.9	Passed 12/72	llama.cppextraction	2026-10-11
toolcallk2-3.7b-b11552-nocoh	100.0%	94.0–100.0	Passed 60/60	llama.cppextraction	2026-10-11
toolsk2-3.7b-b11552-nocoh	86.7%	62.1–96.3	Passed 13/15	llama.cppextraction	2026-10-11

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_0.gguf` (made as `_q40test/K2-Horizon-3.7B-Q4_K_M__Q4_0.gguf`)
Format	Q4_0, 3,022,781,824 B
Where from	made here 2026-09-11 by requantising the Q4_K_M (`--allow-requantize`)
Status	on disk; live DISPATCH seat
Hugging Face	none (no Q4_0 in the IFM repo)

How it was made

Input: models/k2-horizon-3.7b/K2-Horizon-3.7B-Q4_K_M.gguf (see K2-Horizon-3.7B Q4_K_M (quantised here from IFM's BF16)), itself made from the BF16 — this file is rounded twice (BF16 → Q4_K_M → Q4_0).
Command:

llama.cpp-k2horizon/build/bin/llama-quantize --allow-requantize \
K2-Horizon-3.7B-Q4_K_M.gguf models/_q40test/K2-Horizon-3.7B-Q4_K_M__Q4_0.gguf Q4_0

Output: "quant size = 2871.82 MiB (4.76 BPW)". Upstream llama-quantize had failed one minute earlier with unknown model architecture: 'k2-horizon'.

Tool: k2horizon fork @ 35999d1 (build 20). No imatrix.
Date: 2026-09-11; file mtime 14:21:42 BST = 13:21:42Z (stat, 2026-10-23).
Size: 3,022,781,824 B (stat 2026-10-23). GGUF general.file_type = 2 (Q4_0), general.name = Checkpoint_0010000. sha256 not on record.
Who asked / why: part of the Q4_0 sweep after Alan's "we have a lever, chase it, measure it, prove it".
Renamed 2026-09-11 15:51Z to k2-horizon-3.7b/K2-Horizon-3.7B-Q4_0.gguf under Alan's "you can do 1 - 5".
Correction: a later working session says this file "was made locally from the BF16". The command shows it was made from the Q4_K_M.

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-b11552-nocoh`	2026-10-11 10:53Z–10:54Z	upstream llama.cpp b11552 `23b0202a1` (`llama.cpp-b11552`)	`-c 65536 -np 1` / `-ctk q4_0 -ctv q4_0` / `-fa on` / `--jinja` yes / `deepseek`; `-ngl 999`	extraction, `--no-coherence`	upstream quality gate before moving the seat off the fork; the coherence precondition refused the run, and also refused the live fork seat, so it was bypassed
`k2-3.7b-fork-control-nocoh`	2026-10-11 10:55Z	k2horizon fork b20 `35999d1` (launchd plist)	`-c 65536 -np 1` / q4_0 KV / `-fa on` / `--jinja` yes / `deepseek`	extraction, `--no-coherence`	same-day control on the live seat after the 15.3% nullschema
`k2-3.7b-b11552-noreasonfmt`	2026-10-11 10:56Z	upstream llama.cpp b11552 `23b0202a1` (`llama.cpp-b11552`)	as `-nocoh` but no `--reasoning-format`	extraction, `--no-coherence`	tests whether the reasoning-format flag caused the drop

Manifests record model_path = this file and n_ctx 65536, n_slots 1 for all three (result manifests, 2026-10-23).

Seats held:

2026-09-11 15:53Z — fleet seat, -c 8192 -np 2 -fa on -ctk q8_0 -ctv q8_0 -ngl 99 --reasoning-format deepseek (no --jinja); Hermes aux title_generation, profile_describer, then mcp, skills_hub from 16:09Z.
2026-09-12 10:04Z — models/tool-seat-k2-3.7b.sh: --jinja --reasoning-format deepseek -c 32768 -np 1 -fa on -ngl 999, f16 KV; -c 8192 from 10:07Z. Unrouted from 10:09Z.
2026-09-19 — fleetview k2-3.7b → the seat, reasoning_effort: "none" pinned because "low" leaks an orphaned </ifm|think>.
2026-09-19 16:37Z → present — -c 65536 -np 1 -ctk q4_0 -ctv q4_0 -fa on --jinja --reasoning-format deepseek -ngl 999 --alias k2-3.7b; under launchd since 2026-09-30; the live process was started 2026-10-11 14:48 local (ps, 2026-10-23). DISPATCH verb (fleet-verbs).

Decisions

Made to price the Q4_0 speed lever on a dense K2.
Seated without a quality run 2026-09-11 15:53Z. The 36B's Q4_0 got a five-suite board first; this file got none until 2026-10-11, and that run was a build gate, not a seat decision.
Kept as the tool caller 2026-09-18 ("That's your tool caller settled"); permanent DISPATCH worker from 2026-09-20.
Not moved to upstream yet: b11552 passed for dispatch; the plist still names the fork on 2026-10-23.
Not published. No reason is on record. It is a double-rounded requant with an unresolved nullschema question, so a single-pass build would be the candidate.

What went wrong

Coherence precondition fails on this model (fork and upstream alike): reasoning text lands in content for one-word probes ("The user wants me to reply with exac…"). The October rows carry coherence: skipped.
Seated on speed alone (above).
Wrong runtime tried 2026-09-18: .local/bin/llama-server (upstream b10893) failed with unknown model architecture.
KV silently f16 in the seat script until 2026-09-19.
No --jinja on the 2026-09-11 fleet launch; no run-log label used that server.
Refuses general chat (the "Bambu Lab" refusal). Label "TOOL SEAT — refuses general chat".
Zero production requests in the seat log 29 Sep – 9 Oct.
Partial-run check: all three nullschema files have 72 rows; toolcall 60, tools 15 (counted 2026-10-23).

What happened to it

Requantised 2026-09-11 13:21Z → renamed 15:51Z → seat 2026-09-11 → tool router / DISPATCH 2026-09-12 → 2026-09-20 → launchd 2026-09-30 → upstream gate 2026-10-11 → on disk and serving at models/k2-horizon-3.7b/K2-Horizon-3.7B-Q4_0.gguf (3,022,781,824 B, checked 2026-10-23). Not on HF. Re-creatable only via BF16 → Q4_K_M → Q4_0 with the commands in K2-Horizon-3.7B Q4_K_M (quantised here from IFM's BF16) and above; a single-pass BF16 → Q4_0 would be a different artefact.