FreshContextA context lab · Fife
Drawing
02 · Benchmarks
Sheet
2 of 3
Rev
A · 2026-10-01

Ternary-Bonsai-2-27B · file← Ternary-Bonsai-2-27B

Bonsai-2-27B DFlash2 drafter Q8_0

A community draft model meant to speed it up. It wouldn't load on the Mac; deleted the same day. ProCreations.

Size
2.0 GB
Status
Deleted 18 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.

FileBonsai-2-27B-DFlash2-Q8_0.gguf (~2 GB)
Formatdrafter GGUF
Where fromdownloaded 2026-09-18 from ProCreations/Ternary-Bonsai-2-27B-DFlash2
Statusdeleted 2026-09-18 11:31Z
Hugging Facehttps://huggingface.co/ProCreations/Ternary-Bonsai-2-27B-DFlash2

Where it came from

  • Source: ProCreations/Ternary-Bonsai-2-27B-DFlash2, file Bonsai-2-27B-DFlash2-Q8_0.gguf (2.06 GB on the HF listing; 1.9G on disk). The repo also ships model.safetensors (3.85 GB), a prebuilt llama-bonsai-dflash2-linux-cuda13.3-sm120 runtime and a source tarball. The card calls it "an independent experimental adaptation by ProCreations, not an official Prism release", adapted from the Qwen3.8-27B DFlash2 head to the frozen Ternary Bonsai 2 27B PQ2_0 target (listing).
  • Downloaded: 2026-09-18 07:21Z, complete by 07:22Z, into models/bonsai2-dflash2/. Revision and sha256: not on record. This is a draft model, not Bonsai-2 weights, listed here because the convention asks for every file held.

How it was served

labeldates (UTC)binary and versionctx / -np / KV / -fa / --jinja / --reasoning-formatharness profilewhy this label exists
no run-log label; never loaded2026-09-18 07:22Z to 07:34Z (three attempts)llama.cpp-prism fork (b1-6ac5eb0)-m <PQ2_0> -md <this file> -c 8192 -np 1 -ngl 99 -ngld 99 --jinja --no-warmup; the drafter alone with -c 2048 -ngl 99 --no-warmupnoneTest the vendor-adjacent 3x speculative-decoding claim for Bonsai-2 on Metal.

Seats held: none.

Decisions

  • Pulled because the prism build "supports draft models" and the file made the 3x claim testable. The card's 1.038x undercut the claim before any run.
  • Pairing target per the card: PQ2_0, not PTQ1_0.
  • Deleted 2026-09-18 11:31Z with the rest of the chad detour (Alan's answer "All of it — 40 GB (Recommended)", rm -rf). The HF hub cache keeps a 12 KB shell models--ProCreations--Ternary-Bonsai-2-27B-DFlash2 (read 2026-10-23).
  • Published? No.

What went wrong

  • Silent probe: the first two checks printed nothing because macOS has no timeout; the assistant redid them.
  • Lab note disagreement: a lab note says no drafter exists for Bonsai 2; two third-party DFlash2 drafters for it (this one and the nathansutton pack's) were on Hugging Face that morning. The note means PrismML ships none.
  • No RESTORE entry; standing run-log checks do not apply.

What happened to it

Downloaded 2026-09-18 → three failed loads → deleted 2026-09-18 11:31Z. Re-downloadable from ProCreations/Ternary-Bonsai-2-27B-DFlash2.