Muse-Glimmer-30B · file← Muse-Glimmer-30B

kquant-dynamic Q4_K

Meta's own GGUF, behind every August llama.cpp measurement. Deleted twice. Meta GGUF.

Size: 19.6 GB
Status: Deleted 26 Aug 2026
Scored runs: 0

ResultsNever scored

This file never completed a scored run. Its record below says why.

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	`muse-glimmer-30B-kquant-dynamic.gguf` (19,640,844,288 B)
Format	GGUF Q4_K ("Q4_K - Medium" in llama-bench)
Where from	downloaded ~2026-08-11; re-fetched 2026-09-08 as `Muse-Glimmer-30B-KQuant-Dynamic-Q4_K_XL.gguf`
Status	deleted 2026-08-26; re-fetched copy deleted 2026-09-08
Hugging Face	meta-models/Muse-Glimmer-30B-GGUF

Where it came from

First copy (August): meta-models/Muse-Glimmer-30B-GGUF, file muse-glimmer-30B-kquant-dynamic.gguf, fetched by models/muse-glimmer-30b/get.sh with pget.sh (6 range requests) beside mmproj-kquant.gguf and dflash-kquant.gguf. Date: not on record; between the model's release (2026-08-09) and the 2026-08-11 benchmark note. All August weights were "SHA-256 verified against HF LFS OIDs"; the digests were not kept. Who asked: not on record.
Size: 19,640,844,288 bytes (llama-bench model_size, eval/results_fleet_bench.json); a lab note gives 18.29 GiB, Q4_K_M. llama-bench labels it "Q4_K - Medium".
Second copy (2026-09-08): hf download meta-models/Muse-Glimmer-30B-GGUF Muse-Glimmer-30B-KQuant-Dynamic-Q4_K_XL.gguf --local-dir muse-glimmer-30b-meta, queued 10:03Z to get a target from the same publisher as the draft, landed ~13:00Z after a stalled Wi-Fi link. The listing gave 18.3 GiB. (reconstructed) that this is the same upstream file under a new name, because the byte size matches to the precision available; the hashes were not compared.
Companions: Muse-Glimmer 30B vision projector (mmproj-kquant, Meta), Muse-Glimmer 30B DFlash draft (Meta, 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
none in the run log	August 2026 (lab notes); llama-bench 2026-08-23 13:14Z; DFlash retest 2026-09-08 14:24Z	b10369 (`6e62ba538`) in August; b10618 for the retest	recommended invocation: `llama-server -m …kquant-dynamic.gguf -ngl 99 --spec-type draft-dflash -md …dflash-kquant.gguf --mmproj …mmproj-kquant.gguf --jinja` (lab notes); retest: `-c 8192 --jinja -ngl 999 --alias mg` plus `--model-draft … --spec-draft-n-max 15 --spec-draft-n-min 1` for the speculative arm	none (pre-run-log; no harness suite)	the first Muse-Glimmer record and the preset behind the 2026-08-23 fleet table

Seats held: router preset [muse-glimmer-30b] (model, mmproj, spec-type = draft-dflash, model-draft) in fleet-models.ini until commented out on 2026-08-26 10:03Z; never a standing seat.

Decisions

Downloaded as Meta's first-party set: model, projector and draft (get.sh header).
Kept as the router preset through 2026-08-26, then deleted with the rest of the August Muse files on Alan's "nuke the q5 and glimmer"; the assistant noted the weights deleted were 21 GB and the benchmark results survive in four result files.
Re-fetched on 2026-09-08 only as the matched DFlash target, then deleted at 16:14Z on Alan's "sure nuke them for now" to free space for a REAP conversion.
Not published.

What went wrong

Unmatched DFlash test (2026-09-08 morning) paired unsloth's target with Meta's draft; the assistant blamed the pairing ("I mixed repositories."); the matched retest disproved that (acceptance 11.1% against 10.8%).
Wi-Fi stall. The re-fetch crawled at 22-38 Mbps with a 636 ms average gateway RTT and the downloader needed xet disabled to recover. Cost: not quantified; the matched retest waited on this download and ran at 14:24Z.
Contended baseline. The retest's 5.7 tok/s baseline is 40% below the 9.6-9.95 seen elsewhere with the same weights; swap use was 5.7 GB and downloads were running. Finding Speculative decoding did not help, by either method tested quotes that baseline without a caveat.

What happened to it

Downloaded ~2026-08-11 → benched 2026-08-11..23 → preset → deleted 2026-08-26 10:03Z (the muse-glimmer-30b directory, 21 GB with the projector and draft) → re-fetched 2026-09-08 into the muse-glimmer-30b-meta folder (24.5 GB on disk incl. cache) → deleted 2026-09-08 16:14Z (rm -rf, 24 GB). No RESTORE entry for either deletion (both predate the RESTORE pattern's use for model dirs; RESTORE-deleted-20260908.sh has no Muse lines, checked 2026-10-23). Re-downloadable from meta-models/Muse-Glimmer-30B-GGUF.