Ornith 1.5 35B-A3B · file← Ornith 1.5 35B-A3B
Q5_K_M
The first copy we held, replaced by the faster Q6_K. Deleted unscored. Downloaded, deleted unscored.
- Size
- 25.4 GB
- Status
- Deleted 26 Aug 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 | Ornith-1.5-35B-Q5_K_M.gguf |
| Format | Q5_K_M, 25.35 GB |
| Where from | downloaded 2026-08-20 (ornith-dl.log); repo not in the log |
| Status | deleted 2026-08-26 10:03Z |
| Hugging Face | https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B-GGUF (filename match) |
Where it came from
- File:
models/ornith-1.5-35b/Ornith-1.5-35B-Q5_K_M.gguf, 25.35 GB, plusmmproj-Ornith-1.5-35B-BF16.gguf0.90 GB; "26.25GB in 1500s = 140 Mbps" (models/ornith-dl.log, mtime 2026-08-20 21:45 local). - Repo: not in the log. The filename matches
ornith-ai/Ornith-1.5-35B-A3B-GGUF, which listedOrnith-1.5-35B-Q5_K_M.ggufat 23.61 GiB on 2026-09-07. A name match is not proof of the source. - Revision, sha256: not on record (the download script is not on record).
- Date: 2026-08-20. Who asked: not on record; no working session survives for 20 Aug.
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
| — (no run-log label) | — | — | — | — | never served to the locked harness |
Off the run log it was served through the router preset [ornith-1.5-35b] (262,144 ctx, q8_0 KV, reasoning-budget = 8000), briefly alongside a separate Q6_K entry on 2026-08-21 (fleet-models.ini). Seat: resident default on 2026-08-20, replaced by Q6_K the same week.
Decisions
- Downloaded 2026-08-20 as the qwen3.6-35b-a3b replacement (family doc).
- Replaced by Q6_K: "Q6_K over Q5_K_M because it measured FASTER as well as more precise" (
fleet-models.ini). - Deleted 2026-08-26 10:03Z. Alan: "Ornith still may be useful for coding, keep it, nuke the q5 and glimmer"; the assistant had said "Q5 has no remaining purpose".
What went wrong
- Duplicate presets crash (2026-08-21): Q5 and Q6 registered as separate models;
--models-max 2loaded both, 25.35 + 29.21 = 54.6 GB against a ~51.8 GB working set, and llama.cpp died with "Compute error. off = 0, n_batch = 2048, ret = -3". Spumco hit it by using the model name from a lab note (fleet-models.ini; a lab note §6). Fix: one preset name per model. - Unexplained speed inversion: 9% slower than the larger Q6_K. Two attempts to model the per-token read gave answers over 100% of the memory ceiling; the later explanation was that a 3B-active MoE is not bandwidth-bound, so the dequantisation cost dominates.
What happened to it
Downloaded 2026-08-20 → resident default 2026-08-20 → replaced by Q6_K by 2026-08-21 → llama-bench 2026-08-23 → deleted 2026-08-26 10:03Z. No RESTORE entry. Re-downloadable, if the filename match holds, from https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B-GGUF.