Devstral Small 2 24B · file← Devstral Small 2 24B
Q6_K
A larger, more precise copy, downloaded as a quality reference but never run. Kept over the scored file, then deleted too. Unsloth.
- Size
- 19.3 GB
- Status
- Deleted 11 Oct 2026
- Scored runs
- 0
This file never completed a scored run. Its record below says why.
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 | Devstral-Small-2-24B-Instruct-2512-Q6_K.gguf |
| Format | Q6_K GGUF, 19,346,476,192 B |
| Where from | downloaded 2026-08-24, sha256 verified |
| Status | deleted 2026-10-11 (Tier B sweep) |
| Hugging Face | same repo |
Where it came from
- Repo / file:
unsloth/Devstral-Small-2-24B-Instruct-2512-GGUF/Devstral-Small-2-24B-Instruct-2512-Q6_K.gguf. - Size: 19,346,476,192 B (HF tree and
ls -laon 2026-08-24, 19.35 GB in the 2026-09-12 inventory). - sha256:
b04b3e19730d7a1e19530f40b947b69c028b090cb7c58c1515cf1fc2ece5f821, checked by a one-off script. - Revision: not on record.
- Date: 2026-08-24; file mtime 17:43 BST (16:43Z).
- Why: "Q6_K is worth having as a quality reference for extraction work, where there's no wall clock" (assistant). Task #14: "Q6_K (19.4GB, extraction quality reference)".
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
| — (none) | — | — | — | — | never served: no script, run-log row, preset or seat references this file |
The only "use" on record is header-only. seat_size.py read its GGUF header on 2026-09-14 and 2026-09-18 to estimate seat cost: dense, 19.3 G weights, 160 KB/token KV, 2.68 GB at 65,536 with q4_0 KV, 22.0 G per seat. It was also listed (not run) as a dense calibration candidate on 2026-09-11.
Decisions
- Downloaded as an extraction-quality reference. The comparison was never run; no reason is on record.
- Kept over the scored UD-Q4_K_XL on 2026-09-07 ("rm devstral UD-Q4_K_XL (kept Q6_K)"). This followed the "keeping the better one of each" rule, applied by bit-width.
- Marked KEEP on 2026-09-16 as "tested, no faster build on disk". It was not itself tested; the label's runs were on the deleted UD-Q4_K_XL.
- Deleted 2026-10-11 in the Tier B sweep under Alan's "nuke tiers a & b" (a lab note).
What went wrong
- Credited with another file's scores.
RESTORE-deleted-20261011-tiersAB.shand a lab note describe this 18 GB dir as "Benched 2026-09-08: grounding 79.2, nullschema 62.5, screening 100.0 (b10618)". Those runs were on 2026-08-25, on the UD-Q4_K_XL, which had already been deleted on 2026-09-07. This file was never benched.
What happened to it
Downloaded 2026-08-24 → never served → kept as the "better" duplicate 2026-09-07 → deleted 2026-10-11 (RESTORE-deleted-20261011-tiersAB.sh entry exists but has no hf download line). Re-downloadable from unsloth/Devstral-Small-2-24B-Instruct-2512-GGUF; verify against sha256 b04b3e19…2fe5f821.