Ternary-Bonsai-27B · file← Ternary-Bonsai-27B
MTP TQ2_0
A community copy with a built-in draft layer. Probed for speed and quality, never put through our tests; deleted. Vinpix community GGUF.
- Size
- 8.8 GB
- Status
- Deleted 11 Oct 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 | Ternary-Bonsai-27B-MTP-TQ2_0.gguf, ~8.79 GB |
| Format | community TQ2_0 + MTP head |
| Where from | downloaded 2026-09-11 from vinpix/Ternary-Bonsai-27B-Stock-MTP-GGUF |
| Status | deleted 2026-10-11 |
| Hugging Face | https://huggingface.co/vinpix/Ternary-Bonsai-27B-Stock-MTP-GGUF |
Where it came from
- Source:
vinpix/Ternary-Bonsai-27B-Stock-MTP-GGUF, fileTernary-Bonsai-27B-MTP-TQ2_0.gguf, a community conversion, not prism-ml's, pulled intomodels/ternary-bonsai-27b-mtp/on 2026-09-11 (started 19:45Z) byhfpull.shfive range streams. - Size: about 8.79 GB (
8.2Gdu; 8.78 target at the time of the pull; exact bytes not recorded). Revision and sha256: not on record. - Header facts (read 2026-09-14): arch
qwen35, 866 tensors, 65 blocks withnextn_predict_layers = 1(the MTP head is block 64, inside the file), vocabulary 248,320, 28.59 B parameters, 2.46 bits per weight,file_type 37(TQ2_0): 480 tensors / 24.27 B parameters in TQ2_0 and 27 tensors / 4.33 B parameters in Q2_K, Q4_K, Q5_K and Q6_K, against prism's own card claim of "no high-precision escape hatches" for its builds (tensor census, load log). - Why this one: the assistant pulled it as the MTP self-speculation test, together with the 8B ternary, after Alan asked for more ternary models.
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 20:06Z functional check; 20:59Z speed; 21:11Z perplexity | stock .local/bin/llama-cli / llama-perplexity on branch fc/diffusion-confidence @ 63cbb9227 | -c 2048 -ngl 99 -n 64 --temp 0 -st -v --spec-type draft-mtp --spec-draft-n-max 3 | none | Probe of self-speculation and of what a trained-ternary container does to speed and perplexity. Never benched on the gauntlet. |
Seats held: none.
Decisions
- Alan, 2026-09-11 19:38Z: "Read all of the docs for squish, also look for more ternary models online, bonsai was impressive". He then asked "Can we ternary models ourselves or do you need the actual training data", and at 20:06Z: "Nah ignore ternary for now". The downloads finished anyway and were parked, not deleted.
- Alan, 20:58Z: "No take the 27b-mtp for sure, I meant making our own"; the assistant: keeping it, not deleting. (reconstructed) The assistant had asked whether to drop the parked ternary downloads; Alan meant keep this file and stop the idea of ternarising our own models.
- What it settled: a trained ternary container is exact in TQ2_0 and can be both faster than Q4_0 per byte and slower per parameter; post-hoc ternarisation (the assistant's 9B experiment, 44 tensors ternarised to 31.1% zeros) destroyed the model (TQ2_0 is fast and lossless to pack, but ternary weights cannot be made post hoc).
- Not benched: on 2026-09-14 the assistant re-offered the TQ2_0 build as an alternative to Q2_g64 ("the MTP/TQ2_0 build is the one your finding cites"); Alan chose neither and sent the work to the prism fork.
- Published? Only through the finding above.
What went wrong
- Not prism's artefact: 4.33 B parameters sit in K-quants, so 8.79 GB and 2.46 bpw overstate what a pure ternary 27B costs. Prism's card claims no high-precision escape hatches for its own builds; this file has them.
- A scoring-free file: RESTORE-deleted-20261011-tiersAB.sh says "NEVER benched under this name" — correct; there are no run-log rows. Stale build stamps / partial runs / missing flags: not applicable, there are no run-log rows. The CLI runs had no
--jinja. - Perplexity comparison incomplete: the Qwen3.8-27B Q4_0 perplexity arm timed out at two minutes, so no 27B Q4_0 perplexity exists on this corpus.
What happened to it
Downloaded 2026-09-11 → probed 2026-09-11 → parked → deleted 2026-10-11 (ternary-bonsai-27b-mtp, 8.2 GB, second sweep; RESTORE-deleted-20261011-tiersAB.sh). Re-downloadable from vinpix/Ternary-Bonsai-27B-Stock-MTP-GGUF.