Muse-Glimmer-30B · file← Muse-Glimmer-30B
UD-Q4_K_XL
Unsloth's GGUF, and the only file with scores. Deleted on 9 September. Unsloth GGUF.
- Size
- 15.9 GB
- Status
- Deleted 9 Sep 2026
- Scored runs
- 2
- Smallest Mac
- 32 GB measured, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| groundingmuse-glimmer-30bunenforced | 85.6% | 78.4–90.7 | Passed 107/125 | llama.cppextraction | 2026-09-08 |
| screeningmuse-glimmer-30b | 86.7% | 75.8–93.1 | Passed 52/60 | llama.cppextraction | 2026-09-08 |
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.
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-UD-Q4_K_XL.gguf (15,878,222,368 B) |
| Format | GGUF UD-Q4_K_XL |
| Where from | downloaded 2026-09-07/08 |
| Status | deleted 2026-09-09 |
| Hugging Face | unsloth/Muse-Glimmer-30B-GGUF |
Where it came from
- Repo and file:
unsloth/Muse-Glimmer-30B-GGUF,Muse-Glimmer-30B-UD-Q4_K_XL.gguf. Command (queued 2026-09-07 21:56Z, re-issued 23:03Z):
hf download unsloth/Muse-Glimmer-30B-GGUF Muse-Glimmer-30B-UD-Q4_K_XL.gguf --local-dir muse-glimmer-30b-gguf (run from models, with HF_HUB_ENABLE_HF_TRANSFER=1 HF_XET_HIGH_PERFORMANCE=1).
- Landed: between 2026-09-08 05:55Z (the file was named in the finish chain) and 07:30Z (listed complete).
- Who asked: the 2026-09-07 re-examination (see the family doc). Alan did not name this file; he asked for "muse glimmer on mlx". Choosing the unsloth text file plus an Unsloth Q8_0 projector was the assistant's, because the assistant found the GGUF repos ship projectors.
- Companions: Muse-Glimmer 30B vision projector (mmproj Q8_0, unsloth), 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 |
|---|---|---|---|---|---|
muse-glimmer-30b (screening) | 2026-09-08, started ~07:12Z (script written 05:55Z), result 07:46Z (run-log 08:46:40 BST) | llama.cpp b10618 (eb25b7263) | 16384 / default / f16 / default / yes / not set; -ngl 999; no --mmproj, no draft | extraction (thinking off, 2048 max tokens, reasoning budget 0) | first locked-harness scores for a model that had only been ruled out on speed |
muse-glimmer-30b (grounding) | started after screening; killed 09:17Z at 83 of 125 rows and resumed by a one-off script (same flags, resume on), result 10:12Z (run-log 11:12:06 BST) | b10618 | 16384 / default / f16 / default / yes / not set | extraction | same sweep; the kill was to drop the redundant three-suite DFlash run |
The label is a plain llama.cpp label; the speculative runs used the separate throwaway aliases muse-glimmer-30b-spec and mg-spec (no run-log entries; no result file for muse-glimmer-30b-spec exists, so the planned 20-case screening sample on the spec server left nothing).
Seats held: none. Not in fleet-models.ini at any time after 2026-08-26.
Decisions
- Pulled for the vision path and DFlash draft after the old speed verdict turned out to be unscored (family doc).
- Kept only until the three suites and the DFlash probe were done, then deleted as redundant on Alan's "muse glimmer guff can go". Reasoning given to Alan: no live references, MLX copy kept as the vision route.
What went wrong
- Mid-run kill and resume, 2026-09-08 09:17Z. Grounding was killed at 83/125 and resumed; the manifest keeps both segments. Cost: none beyond the restart; no rows lost or duplicated (125 rows).
- Harness edited in place, 2026-09-08 ~10:13-10:25Z. To stop MLX servers that ignore
response_formatfrom being scored on schema suites, the assistant added a guard toharness/runner.py; its first placement raisedUnboundLocalError(beused before construction), which is why the planned nullschema step for this model never ran, and a later edit changed the sentinel probe to 512 tokens. These were in-place edits to a harness the project treats as locked; the repo history starts on 2026-09-30, so there is no diff for them. The probe was extended on 2026-10-21 by PR #9 (commit6c41dca). - Contended DFlash baseline. Test 1 ran beside the grounding suite and downloads; the 2.8 vs 9.6 tok/s ratio is valid, the absolute figures are not.
- Standing checks. Partial runs: both result files equal their suite sizes (60 and 125). Stale build stamps: run-log b10618 matches the 2026-08-25 rebuild and the 2026-09-07
eb25b7263check. Missing flags:--jinjapresent;--reasoning-formatnot set for a reasoning model; whether the extraction profile's "thinking off" took effect on this template is not on record.
What happened to it
Downloaded 2026-09-07/08 → screening + grounding 2026-09-08 → DFlash probes 2026-09-08 → deleted 2026-09-09 14:25Z (rm -rf muse-glimmer-30b-gguf, 18 GB reclaimed). No RESTORE entry: RESTORE-deleted-20260909.sh has no Muse-Glimmer lines (checked 2026-10-23). Re-downloadable from unsloth/Muse-Glimmer-30B-GGUF (the run-log results survive).