Ternary-Bonsai-2-27B · file← Ternary-Bonsai-2-27B
Ternary-Bonsai-2 + DFlash2 MLX pack
A community MLX copy with a built-in draft model. It never loaded; deleted the same day. Nathansutton.
- Size
- 8.2 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.
| File | pack in models/qwen38-bonsai2-mlx (8.2 GB) |
| Format | MLX + bundled drafter |
| Where from | downloaded 2026-09-18 from nathansutton/Qwen3.8-27B-Ternary-Bonsai-2-DFlash2-MLX |
| Status | deleted 2026-09-18 11:31Z |
| Hugging Face | https://huggingface.co/nathansutton/Qwen3.8-27B-Ternary-Bonsai-2-DFlash2-MLX |
Where it came from
- Source:
nathansutton/Qwen3.8-27B-Ternary-Bonsai-2-DFlash2-MLX, a third-party MLX pack:model.safetensors7.67 GB,dflash/model.safetensors1.15 GB,hadamard.json, tagsprism_hadamard_qwen35,ternary,speculative-decoding; card headerbase_model: Qwen/Qwen3.8-27B. The card warns that an ordinary MLX affine loader finds tensors of the right shapes, skips the Hadamard transforms and "produces plausible garbage without raising". - Downloaded: 2026-09-18, started 07:44Z, complete by 08:06Z,
snapshot_downloadintomodels/qwen38-bonsai2-mlx/(8.2 GB on disk). Revision, sha256 and exact bytes: not on record. - Why this one: the "64 t/s" claim Alan quoted came from the pack's card paired with the
chadcoding agent.
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 loaded | 2026-09-18 08:07Z, one attempt | chad-bench from chad-code 2.1.0 (editable install of main @ a4d6588, mlx 0.32.2, mlx-lm 0.31.3, transformers 5.12.1) with CHAD_MODEL pointing at the pack | n/a | none | To measure the drafter claim on this M1 Max. Failed at load: Model type prism_hadamard_qwen35 not supported; chad 2.1.0 and PyPI's latest have no such loader and no branch or PR mentions prism, hadamard or ternary. |
Seats held: none.
Decisions
- Alan asked for the M4 Pro setup prompt and then pulled the thread: "chad is a coding harness, not an llm how the heck did we end up installing it?", "it doesn't even serve over http", "we need to can it, it's not what we need".
- Deleted 2026-09-18 11:31Z after the assistant asked what to clean up; Alan's selections were "All of it — 40 GB (Recommended)", keeping the gemma4-26b MLX and the chad clone and venv. The
rm -rfnamed this directory withqwen38-chad-default,bonsai2-dflash2andqwen38-dflash-patched. - Published? No.
What went wrong
- Plausible-garbage trap avoided: the pack would have loaded as garbage in a stock affine loader; chad refused it by name instead.
- Wrong tool, found late: chad is a TUI/one-shot agent with no HTTP server; the assistant installed it before checking and benchmarked a number that cannot back a fleet seat.
- No RESTORE entry (deleted 2026-09-18, before the RESTORE-script pattern covered it; the 2026-10-11 scripts do not list it). Standing checks (partial runs, stale stamp, flags): no run-log rows.
What happened to it
Downloaded 2026-09-18 → single failed load → deleted 2026-09-18 11:31Z (8.2 GB). Re-downloadable from nathansutton/Qwen3.8-27B-Ternary-Bonsai-2-DFlash2-MLX.