Qwen3.5-9B · file← Qwen3.5-9B
ternarised F16
That copy with its main weights forced to three values. A step on the way; never run. Made here, quantlab intermediate.
- Size
- 17.0 GB
- Status
- Deleted 11 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 | _quantlab/qwen35-9b-tern-f16.gguf |
| Format | F16 holding only {-s, 0, +s} values, 17 GB |
| Where from | made here from the F16 above |
| Status | deleted 2026-09-11 22:28Z |
How it was made
- File:
models/_quantlab/qwen35-9b-tern-f16.gguf(17 GB byls -lh). Made from the F16 above, copying the file and patching tensor data in place at the same offsets. - Command:
ternenv/bin/python ternarise.py models/_quantlab/qwen35-9b-f16.gguf models/_quantlab/qwen35-9b-tern-f16.gguf, in a scratch uv venv (uv venv --python 3.12 ternenv, numpy, pyyaml, tqdm). The rule: BitNet-b1.58 absmean per 256-element block,s = mean(|w|),q = clip(round(w/s), -1, 1), written back asq*sin F16; embeddings, output and norms skipped. - Date / who asked: 2026-09-11 21:01-21:09Z; the assistant, to answer Alan's question about making sub-4-bit models ("Can we ternary models ourselves or do you need the actual training data") and Alan's two questions of 13:31Z and 13:34Z: "also I wonder if q5_0 or q6_0 become the sweet spot?" and "so q6 being divisible by 2 won't help?". The
_quantlabdirectory was created by this command. - sha256: not on record. Exact bytes: not on record (only the rounded sizes shown).
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) | 2026-09-11 21:02-21:11 | .local/bin/llama-cli / llama-perplexity (symlink to llama.cpp, b10893 by reconstruction) | -c 4096 -ngl 99 -n 192 --temp 0 -st (speed); -c 512 -ngl 99 --chunks 12 (PPL) | n/a | alignment sweep: speed per byte |
Seats held: none.
Decisions
- Made to test the claim that
llama-quantize TQ2_0packs already-ternary weights losslessly (the kernel sets the scale to the block's absolute maximum). - Deleted 2026-09-11 22:28:38Z.
What went wrong
- The source was the Q8_0, so "requantising from Q8_0 compounds error -- this measures SPEED/alignment only, not quality. Quality comparison needs a BF16 source" (the BF16 GGUF was never fetched).
- Standing checks: no suite ran, so partial-run, build-stamp and flag checks do not apply.
What happened to it
Made 2026-09-11 -> speed-probed -> deleted 2026-09-11 22:28:38Z. No RESTORE-deleted-*.sh entry exists for it. Re-creatable from the Q8_0 (itself deleted) or from a fresh BF16 download.