Q4_0 GGUF

The smaller copy we tested on tool use. The last file we held. Unsloth.

Size: 17.2 GB
Status: Deleted 11 Oct 2026
Scored runs: 6
Smallest Mac: 32 GB estimate, 8K context

Results6 scored · 95% intervals

Suite	Score	95% interval	Passed	Runtime · profile	Recorded
epistemicsglm-4.7-flash-q40	75.0%	61.2–85.1	Passed 36/48	llama.cppextraction	2026-09-12
nullschemaglm-4.7-flash-q40	44.4%	33.5–55.9	Passed 32/72	llama.cppextraction	2026-09-12
synthesisglm-4.7-flash-q40	60.0%	47.4–71.4	Passed 36/60	llama.cppextraction	2026-09-12
toolcallglm-4.7-flash-q40	38.3%	27.1–51.0	Passed 23/60	llama.cppextraction	2026-09-12
toolscaleglm-4.7-flash-q40	100.0%	91.2–100.0	Passed 40/40	llama.cppextraction	2026-09-12
worldviewglm-4.7-flash-q40	37.5%	24.2–53.0	Passed 15/40	llama.cppextraction	2026-09-12

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	`glm-4.7-flash-gguf/GLM-4.7-Flash-Q4_0.gguf`
Format	GGUF Q4_0, 17,216,676,192 B (HF listing)
Where from	downloaded 2026-09-12 from `unsloth/GLM-4.7-Flash-GGUF`
Status	deleted 2026-10-11 (sweep Tier B)
Hugging Face	as above

Where it came from

Repo / file: unsloth/GLM-4.7-Flash-GGUF / GLM-4.7-Flash-Q4_0.gguf, 17,216,676,192 B (HF listing; "17.22 GB"). Revision and sha256 not on record.
Command:

mkdir -p models/glm-4.7-flash-gguf
$S/hfpull.sh unsloth/GLM-4.7-Flash-GGUF GLM-4.7-Flash-Q4_0.gguf \
models/glm-4.7-flash-gguf/GLM-4.7-Flash-Q4_0.gguf 6

(six range streams.) Started 2026-09-12 10:11Z; 16 G on disk by 10:32Z.

Who asked: Alan, "Been watching videos on GLM 4.7 flash which is a very capable agentic coding model … I’m gonna have to take it for another spin"; "actually think glm 4.7 could be good?". Why Q4_0 is not stated at the time (reconstructed): the previous day's sweep had found Q4_0 the fast format on Metal.
GGUF header (read 2026-09-12 22:38Z): deepseek2, 47 layers, kv_heads 1, head_dim 576 → 99.9 KB/token f16, 25.0 KB/token q4_0.

How it was served

label	dates (UTC)	binary and version	ctx / -np / KV / -fa / --jinja / --reasoning-format	harness profile	why this label exists
`glm-4.7-flash-q40`	2026-09-12 12:06Z–12:14Z	`.local/bin/llama-server` → b10893 `c32d1dabe` (manifest /props `build b10893-c32d1dabe`)	`-c 32768 -ngl 99 --jinja -fa on --alias glm-4.7-flash`; no `-np` given, /props reported `n_slots 4`, `n_ctx 32768`; KV default (`cache_type_k` null); no `--reasoning-format`	extraction (temperature 0, seed 42, thinking off)	"the six suites it has never been run on"

Script: glm_bench.sh, run by the "away tests" chain gated on the driver being idle. The run-log stamp 0.4.0-dev (build 10893, commit c32d1dabe) matches /props, so it is not stale. No seats held.

Decisions

Downloaded 2026-09-12 to test the agentic claim. The bench was held back while Alan was using our agents: "deliberately not queued — running it would contend with your session".
Not seated as coder 2026-09-12 22:53Z. Ornith-35B-A3B was chosen on toolcall 80.0 vs 38.3 and decode 55.3 vs 50.2. Alan had wanted it in DSH/opencode. Whether he tried it there is not on record.
Kept 2026-09-16 as the family's surviving build at 16 GB, after Alan's "Does that leave another GLM flash".
Deleted 2026-10-11 in sweep Tier B, "benched-once, abandoned" (a lab note). Alan: "nuke tiers a & b".
Not published (upstream file).

What went wrong

Wrong provenance in the sweep record. The sweep's restore script and a lab note say "glm-4.7-flash-gguf (16G): never benched as this GGUF build". It was benched as glm-4.7-flash-q40, 6 suites, 2026-09-12.
A RESTORE line that would pull the whole repo. The entry is hf download unsloth/GLM-4.7-Flash-GGUF, with no filename or --include.
-np left to the default: four slots at 32k total context, against the Q8_0 arm's single slot. Not a reported cause of any score, but the configuration differs from the Q8_0 rows it is compared with.

What happened to it

Downloaded 2026-09-12 → benched 2026-09-12 → candidate, no seat → deleted 2026-10-11. Re-downloadable: hf download unsloth/GLM-4.7-Flash-GGUF GLM-4.7-Flash-Q4_0.gguf --local-dir models/glm-4.7-flash-gguf.