Ornith 1.5 35B-A3B · file← Ornith 1.5 35B-A3B
partial PTQ1_0
The shrink on the middle layers only. Passed the check, badly damaged; deleted. Blocks 10–30 ternary, made here.
- Size
- 17.4 GB
- Status
- Deleted 17 Sep 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-35b-mixed/ornith-35b-partial.gguf |
| Format | PTQ1_0 gate/up on blocks 10–30, Q4_0 elsewhere; 16,600 MiB |
| Where from | made here 2026-09-17 |
| Status | deleted 2026-09-17 13:57Z |
How it was made
- Input:
models/ornith-35b-q8.gguf(variant-ornith-1.5-35b-q8-0) andmodels/ornith-35b.imatrix. - Recipe (
models/ornith-35b-mixed/recipe-partial.tt, still on disk):
blk\.(1[0-9]|2[0-9]|30)\.ffn_gate_exps\.weight=ptq1_0
blk\.(1[0-9]|2[0-9]|30)\.ffn_up_exps\.weight=ptq1_0
ffn_gate_exps\.weight=q4_0
ffn_up_exps\.weight=q4_0
ffn_down_exps\.weight=q4_0- Command:
cd llama.cpp-prism
build/bin/llama-quantize --allow-requantize --imatrix models/ornith-35b.imatrix \
--tensor-type-file models/ornith-35b-mixed/recipe-partial.tt \
models/ornith-35b-q8.gguf models/ornith-35b-mixed/ornith-35b-partial.gguf Q8_0 8Base type Q8_0, so the router (ffn_gate_inp), SSM and attention tensors stay Q8_0 (RECIPE.md).
- Tool: prism fork
6ac5eb0(build 1). Output: 16,600.46 MiB, 3.92 BPW, 42 ternary tensors as expected. sha256 not on record. - Date: 2026-09-17, 13:30–13:39Z. Who asked: Alan, "but yes try it on ornith, why not". Design notes in
RECIPE.md(why gate/up only, why down stays Q4_0, why the calibration set excludes suite prompts).
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 models/ornith-35b-mixed/ornith-35b-partial.gguf -c 8192 -np 1 -ngl 99 --no-warmup for the coherence gate.
Decisions
- Built to test whether selective ternary on routed-expert gate/up keeps Ornith usable (family doc).
- Rejected: "The shrink thesis is closed and the answer is no. Ornith's floor is the Q4_0 seat".
- Deleted 2026-09-17 13:57Z on Alan's cleanup answer. Not published.
What went wrong
- Gate pass on a damaged model: "the partial-ternary Ornith passed this gate at 7/8 with confidence 0.64 while carrying 2.9x the perplexity of its Q8 reference. The gate separates rubble from usable; it does not rank the usable" (A destroyed model and an incapable one produce the same zero, so the harness now refuses to run without a coherence check, action).
What happened to it
Made 2026-09-17 → coherence gate + perplexity → deleted 2026-09-17 13:57Z. No RESTORE entry. Rebuildable in about a minute from the kept recipe and imatrix once a Q8_0 exists.