Muse-Glimmer-30B · file← Muse-Glimmer-30B
DFlash draft
Meta's small draft model. Faster at one setting, slower at another. Meta, Q4_K_M.
- Size
- 1.5 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 | dflash-kquant.gguf (Aug), dflash-Muse-Glimmer-30B-Q4_K_M.gguf (Sep, 1.5 GB) |
| Format | GGUF DFlash draft |
| Where from | downloaded ~2026-08-11 and 2026-09-08 |
| Status | Aug copy deleted 2026-08-26, Sep copy 2026-09-09 |
| Hugging Face | meta-models/Muse-Glimmer-30B-GGUF |
Where it came from
- August copy:
dflash-kquant.gguffrommeta-models/Muse-Glimmer-30B-GGUFviaget.sh; date and size not on record. - September copy:
hf download meta-models/Muse-Glimmer-30B-GGUF dflash-Muse-Glimmer-30B-Q4_K_M.gguf --local-dir muse-glimmer-30b-gguf; 1.5 GB (ls). Landed by 2026-09-08 07:30Z. - The two names are most likely one upstream file renamed ("kquant" to "Q4_K_M"), judged from the matching 1.5 GB size and the repo's later file list; the bytes were not compared.
- It is a 1.5 GB draft model for speculative decoding trained against Meta's own 30B target with block size 15 (the server clamped
n_max 16to 15).
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 (n=3, n=5); 2026-09-08 09:47-10:02Z (test 1); 2026-09-08 14:24Z (matched retest) | b10369 (August), b10618 | August: --spec-type draft-dflash -md …; test 1: --model-draft … --spec-draft-n-max 16 --spec-draft-n-min 1 -c 8192 --jinja -ngl 999; retest: --spec-draft-n-max 15 --spec-draft-n-min 1 | none | does a published draft make a dense 30B usable? |
Seats held: the August router preset only (model-draft = …/dflash-kquant.gguf).
Decisions
- Pre-registered by the assistant before the September test: "a matched 1.5 GB draft is exactly the lever that could move a dense 30B into usable territory". Refuted at depth 15.
- Alan trimmed the speculative run to a sample.
- Finding Speculative decoding did not help, by either method tested was written from the September result, with the mismatched-pair hypothesis recorded as disproved. It does not mention the August depth-3 gain.
What went wrong
- Silent no-op (August):
--spec-type draft-dflashwithout-mdruns at baseline with no warning (a lab note). - Removed flags (September): test 1 first used
--draft-max 16 --draft-min 1, which the b10618 server no longer accepted; relaunched with--spec-draft-n-max/-min. - Contention: test 1 ran beside the grounding suite; the retest baseline is 40% below the morning's 9.6 tok/s (see the target variant doc).
- Depth mismatch: the finding's "neither method improved decode speed" is untested at depth 3 in September.
- Standing checks: no run-log entries.
What happened to it
August copy deleted 2026-08-26 with the muse-glimmer-30b directory; September copy deleted 2026-09-09 14:25Z with the muse-glimmer-30b-gguf folder. No RESTORE entry. Re-downloadable from meta-models/Muse-Glimmer-30B-GGUF.