Kolibri-1 78B-A3.5B · file← Kolibri-1 78B-A3.5B
Q8_0
A near-full-precision community copy, 83.1 GB, downloaded to make the Q4_0 from. Never scored. Hob-forge two-part GGUF; quantisation source and PPL reference.
- Size
- 83.1 GB
- Status
- On disk
- 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 | Kolibri-1-Q8_0-0000{1,2}-of-00002.gguf (83.1 GB) |
| Format | GGUF Q8_0 |
| Where from | downloaded twice (2026-10-05, deleted; 2026-10-11, on disk) |
| Status | on disk (second copy) |
| Hugging Face | Hob-forge/Kolibri-1-GGUF |
Where it came from
- Source:
Hob-forge/Kolibri-1-GGUF,Q8_0/Kolibri-1-Q8_0-00001-of-00002.ggufand-00002-of-00002.gguf(models/kolibri-1-gguf/download-q8.sh). Revision not on record. - Bytes (stat, 2026-10-23): part 1 = 44,679,459,296; part 2 = 38,455,934,400 (83,135,393,696 total, "83.1 GB"). sha256 not on record; the 2026-10-05 copy was verified against HF sizes.
- Copy 1:
pardl.shwith 6 range parts per file, 2026-10-05, ~10:07Z-10:46Z (download-q8.log: ALL DONE 11:46:53 local). Why: source for the local Q4_0 (Kolibri-1 Q4_0 (made here from the Hob-forge Q8_0, --allow-requantize)), after Alan's "4_0 works better". - Copy 2: the same script re-run 2026-10-11, part 1 landed 13:22Z, part 2 ALL DONE 13:59Z (
download-q8-20261011.log). Why: the assistant proposed building a 2-bit mix and listed re-downloading the faithful 8-bit (83 GB) as step one of the recipe; it called the download "the fallback for the 2-bit route". Alan's prompt for that line of work: "it may be a way of getting the bigger parameter model in the swap as it's the biggest model and the fastest we have".
How it was served
Never served by llama-server. Used as a perplexity reference only (CPU mmap on 2026-10-05 after a Metal out-of-memory error).
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
| — (no run-log label) | 2026-10-05 ~12:00-12:41Z | llama-perplexity from llama.cpp-kolibri/build/bin (b11396, 2e7c58c54 + patch) | -ngl 0 -t 8 -c 2048 -b 512 --chunks 10 | none | Q8_0 reference for the Q4_0 perplexity |
Seats held: none.
Decisions
- Pulled to be quantised, not to be benched. Copy 1 was deleted as soon as the Q4_0 existed. Copy 2 was kept on disk after the 2-bit plan was overtaken by the ready-made Sakura quants (
family-sakura-kolibri-1-365e); the keep decision itself is not on record. - Not published.
What went wrong
- First attempt to use it as a quantisation source failed silently without
--allow-requantize(see Kolibri-1 Q4_0 (made here from the Hob-forge Q8_0, --allow-requantize)). - Metal OOM at 83 GB when used as the perplexity reference.
What happened to it
Downloaded 2026-10-05 -> source for Kolibri-1 Q4_0 (made here from the Hob-forge Q8_0, --allow-requantize) and PPL reference -> deleted 2026-10-05 12:42Z -> re-downloaded 2026-10-11 -> on disk 2026-10-23. No RESTORE entry for the first deletion (it predates the sweep scripts); re-downloadable with models/kolibri-1-gguf/download-q8.sh.