FreshContextA context lab · Fife
Drawing
02 · Benchmarks
Sheet
2 of 3
Rev
A · 2026-10-01

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

MLX 4-bit

Kept as the route to reading images, but never run. Deleted on 11 October. Mlx-community, Sept copy.

Size
18.1 GB
Status
Deleted 11 Oct 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.

Filedir muse-glimmer-30b-mlx-q4 (18.1 GB, 4 shards)
FormatMLX 4-bit
Where fromdownloaded 2026-09-08
Statusdeleted 2026-10-11
Hugging Facemlx-community/Muse-Glimmer-30B-4bit

Where it came from

  • Repo: mlx-community/Muse-Glimmer-30B-4bit (the HF cache holds an empty models--mlx-community--Muse-Glimmer-30B-4bit stub, the local dir is models/muse-glimmer-30b-mlx-q4). Four safetensors shards, 18.1 GB.
  • Command: hf download mlx-community/Muse-Glimmer-30B-4bit --local-dir muse-glimmer-30b-mlx-q4, launched 2026-09-08 09:59Z, stalled at 7.3 GB, restarted with xet disabled, complete by 12:41Z (completion time (reconstructed)).
  • Why this repo: Alan asked for it: "muse glimmer on mlx would be interesting too". The assistant listed the mlx-community options (4/5/6/8-bit, bf16) and took the 4-bit.
  • Size 18.1 GB (du). sha256, revision: not on record. mlx_lm of the day: 0.32.0.
  • The earlier MLX copy is Muse-Glimmer 30B MLX 4-bit (August copy, source not on record); the vision-capable config (vision_config, image_token_id 200092) was checked on this copy.

How it was served

labeldates (UTC)binary and versionctx / -np / KV / -fa / --jinja / --reasoning-formatharness profilewhy this label exists
nonenever servedn/an/an/adownloaded to be run through the suites; the order of 2026-09-08 put the GGUF first and it was never reached

Seats held: none. Listed as UNTESTED in the inventory of 2026-09-12; after the inventory tool was changed to match on name tokens it was listed as tested on 2026-09-16, and the 2026-10-11 sweep note calls it "benched". Both later readings credit this directory with the GGUF's scores (see Incidents; reconstructed that the join was on the family name).

Decisions

  • Kept on 2026-09-09 as the vision route when the GGUF was deleted ("only vision-capable model left").
  • Deleted 2026-10-11 in Tier A ("superseded by measurement") under Alan's "nuke tiers a & b" (a lab note).

What went wrong

  • Mis-attributed benchmark in the sweep record. The sweep note and RESTORE-deleted-20261011-tiersAB.sh say this directory was "benched: grounding 85.6, screening 86.7". Those are the run-log rows of muse-glimmer-30b, the GGUF (Muse-Glimmer 30B UD-Q4_K_XL (unsloth GGUF)). This copy was never run. The sweep audit's own provenance check joined on the family name.
  • Wrong re-download id. The RESTORE entry says hf download velaia/Muse-Glimmer-30B-MLX-Q4; the assistant's check on 2026-10-11 found it returns HTTP 401 and that the HF cache shows the real source is mlx-community/Muse-Glimmer-30B-4bit. The RESTORE file in fleet-roster still carried the wrong id on 2026-10-23.
  • Download stall (Wi-Fi) as described in Muse-Glimmer 30B kquant-dynamic Q4_K (Meta GGUF).
  • Standing checks: no runs, so none apply.

What happened to it

Downloaded 2026-09-08 → never benched → kept 2026-09-09 → deleted 2026-10-11 in RESTORE-deleted-20261011-tiersAB.sh (entry present, repo id wrong; correct source mlx-community/Muse-Glimmer-30B-4bit).