Q8_0 GGUF

The first file, a near-full-precision copy. It ran the first tests and the long-conversation memory test. Unsloth.

Size: 31.8 GB
Status: Deleted 9 Sep 2026
Scored runs: 3
Smallest Mac: 64 GB estimate, 8K context

Results3 scored · 95% intervals

Suite	Score	95% interval	Passed	Runtime · profile	Recorded
groundingglm-4.7-flash-q8	76.0%	67.8–82.6	Passed 95/125	llama.cppextraction	2026-08-26
nullschemaglm-4.7-flash-q8	48.6%	37.4–59.9	Passed 35/72	llama.cppextraction	2026-08-26
screeningglm-4.7-flash-q8	95.0%	86.3–98.3	Passed 57/60	llama.cppextraction	2026-08-26

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/GLM-4.7-Flash-Q8_0.gguf`
Format	GGUF Q8_0, 31,842,799,968 B (HF listing)
Where from	downloaded 2026-08-26 from `unsloth/GLM-4.7-Flash-GGUF`
Status	deleted 2026-09-09 14:08Z
Hugging Face	https://huggingface.co/unsloth/GLM-4.7-Flash-GGUF

Where it came from

Repo / file: unsloth/GLM-4.7-Flash-GGUF / GLM-4.7-Flash-Q8_0.gguf. Revision not on record. The HF listing printed an empty LFS oid, so sha256 is not on record.
Size: 31,842,799,968 B (HF listing); 29.7 GiB on disk.
Command: a one-off script, made by sed from the qwen3.5-122b downloader with the repo swapped to unsloth/GLM-4.7-Flash-GGUF. Six range streams (download log "GLM-4.7-Flash-Q8_0.gguf (30367 MB, 6 streams) 10:58:00 … DONE 11:21:22" local).
When: 2026-08-26 09:58Z–10:21Z.
Who asked / why: Alan, "Are there any interesting 105b range MOE models we can look at?". Chosen as the intact-precision arm.

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-q8` (first start)	2026-08-26 11:13Z–11:28Z	`llama.cpp/build/bin/llama-server`, b10618 `eb25b7263`	`-c 32768 -np 4 -cb -ngl 99 --jinja`; KV default; no `-fa`; no `--reasoning-format`	extraction	the big-MoE gauntlet as first scripted. Killed: Metal died 0.02 s in with 1,276 compute errors. Left two zero-coverage run-log rows
`glm-4.7-flash-q8` (board)	2026-08-26 11:28Z–~12:13Z	same	`-c 32768 -np 1 -cb -ngl 99 --jinja` ("-np 1, NOT 4"); Toolery at `--concurrency 1`	extraction	same label after the slot fix, "so the window matches every other model measured; only the slot count differs"
ctx ladder (no run-log label)	2026-09-08 ~10:2xZ	`llama.cpp/build/bin/llama-server` (b10618 by the run log of that night)	`--n-gpu-layers 99 --ctx-size 8192…131072 --jinja`, KV default (f16)	—	test of the "should fail early" KV prediction

Decisions

Downloaded 2026-08-26 for the big-MoE arm.
Verdict 2026-08-26: "did not vindicate intact-over-compressed" (a lab note §6).
Kept through 2026-09-08 as "only valid grounding/nullschema source, keep", and as the runtime baseline for MLX q8.
Deleted 2026-09-09 14:08Z. Inventory: "glm-4.7-flash / -q8 30+30G both mixed_3_6 covers it". Alan: "ok remove the granite extras and the superseded models". (judgement worth review) mixed_3_6 had only screening on the board, so it did not cover this file's enforced grounding/nullschema.
Not published (upstream file).

What went wrong

Wrong label caught before any result. The sed-made runner first came out as NAME=glm-4.7-flash-iq3. Killed and relaunched as glm-4.7-flash-q8.
False gate failure. The tool-call gate failed on two trailing bytes; the model's call get_weather {"location":"Warsaw"} was clean. Fixed with raw_decode in three runners.
Four slots at Q8 killed Metal. 1,276 compute errors, "backend is in error state", ~200 MB free. The kill left two zero-coverage run-log rows (2026-08-26T12:28:31, screening and nullschema) and ERROR rows reading URLError: Connection refused, which the clean run overwrote. The run log still carries the two rows unmarked.
Residency, not a lock. The next model (Qwen3-Next-80B) died silently during load because this model's 29.7 GB of Metal buffers had not been reclaimed four minutes after its server was killed (a lab note §5).
"47 GB" in a lab note. The note lists this model at 47 GB, which was the directory with the unrun Q4_K_M beside it. The file is 31.8 GB.

What happened to it

Downloaded 2026-08-26 → benched 2026-08-26 → ladder 2026-09-08 → deleted 2026-09-09 14:08Z. No RESTORE entry. Re-downloadable: hf download unsloth/GLM-4.7-Flash-GGUF GLM-4.7-Flash-Q8_0.gguf (also ggml-org/GLM-4.7-Flash-GGUF).