Ornith 1.5 35B-A3B · file← Ornith 1.5 35B-A3B
Q8_0
Our near-full-precision copy, the yardstick for shrinking. Deleted the same day. Made here from BF16.
- Size
- 37.8 GB
- Status
- Deleted 17 Sep 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 | ornith-35b-q8.gguf |
| Format | Q8_0, 35 GB (36,040 MiB) |
| Where from | made here 2026-09-17 from the BF16 (prism convert_hf_to_gguf.py --outtype q8_0) |
| Status | deleted 2026-09-17 13:57Z |
How it was made
- Input:
models/ornith-35b-hf(variant-ornith-1.5-35b-bf16). - Command:
cd llama.cpp-prism
uv run --with numpy --with sentencepiece --with transformers --with torch --with gguf \
python convert_hf_to_gguf.py models/ornith-35b-hf \
--outfile models/ornith-35b-q8.gguf --outtype q8_0- Tool: prism llama.cpp fork (
llama.cpp-prism, HEAD6ac5eb0, reports build 1;git log,--version2026-10-23). The converter handledqwen3_5_moe. - Output: "Model successfully exported", 35G; 36,040.46 MiB, 8.52 BPW (quantize log). sha256 not on record.
- Date: 2026-09-17 13:07–13:16Z. Who asked: Alan (see BF16 doc).
- imatrix built over it:
llama-imatrix -m ornith-35b-q8.gguf -f .cache/mlx-lm/calibration_v5.txt -o models/ornith-35b.imatrix -ngl 99 -c 512 -b 512: final PPL on the calibration text 5.7097, worst expert coverage 99.61%. The imatrix is still on disk (192,223,936 B).
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: prism llama-server -m ornith-35b-q8.gguf -c 8192 -np 1 -ngl 99 --no-warmup --alias ornith-q8 for coherence probes.
Decisions
- Made as the requantisation source for the shrink test (family doc).
- Deleted 2026-09-17 13:57Z on Alan's cleanup answer.
What went wrong
- The gate failed this healthy model first: 24 tokens with thinking on returned empty
contenton all probes (membership 0%). Fixed to thinking off and 512 tokens. - Two "reference" perplexities:
RECIPE.mdcalls 5.7097 the "Q8_0 reference PPL"; that is the imatrix pass over the calibration text. The held-out reference used in the finding is 6.83.
What happened to it
Made 2026-09-17 → coherence + PPL reference → deleted 2026-09-17 13:57Z. No RESTORE entry. Re-creatable from the BF16 with the command above (a re-run is a new artefact unless its hash matches; none was recorded).