Qwen3.8-27B · file← Qwen3.8-27B
MTP draft head, 3-bit MLX
Downloaded as a 3-bit model; it was only a 186 MB draft head. Never run. Lukaskremla; 186 MB, never run.
- Size
- 0.2 GB
- Status
- On disk
- 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 | qwen3.8-27b-mlx-3bit-mtp/model.safetensors |
| Format | MTP draft head only, 185,849,974 B |
| Where from | downloaded 2026-09-07 |
| Status | on disk, never run |
| Hugging Face | https://huggingface.co/lukaskremla/Qwen3.8-27B-MTP-3bit-MLX |
Where it came from
- Repo / file:
lukaskremla/Qwen3.8-27B-MTP-3bit-MLX, onemodel.safetensorsof 185,849,974 B (stat, 2026-10-23), plus tokenizer, template and README. Revision and sha256 not on record.config.json:model_type qwen3_5_mtp,block_size 3, 3-bit affine group 64,vision_config {}. - Date and tool: queued 2026-09-07 19:47Z as the second download behind Qwen3-Next MLX, via
snapshot_downloadin a one-off script intomodels/qwen3.8-27b-mlx-3bit-mtp; file mtimes 20:25Z. At 20:16Z it was still queued. - Identified as a head: 20:32Z. The assistant opened it and found one 186 MB tensor file with
model_type qwen3_5_mtp: 'that's not a model, it's the MTP draft head'.
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
| — (never served) | — | — | — | — | not loadable on its own; the only possible use is as a draft for the 4-bit target, and no such run is on record |
Seats held: none.
Decisions
- Downloaded on the assistant's recommendation, approved implicitly by Alan's wish for speed. Kept through the 2026-09-11 disk sweep: Alan said "keep any q4_0 and drafters", and the 2026-09-11 disk sweep listed it among the 'MTP assets'. Still on disk 2026-10-23.
- Not published.
What went wrong
- Mistaken identity. The assistant had described the repo as 'Qwen3.8-27B-MTP-3bit-MLX ← 3-bit + speculative decoding' and the download as the speed experiment for the dense 27B. It was a 186 MB head with nothing to decode with. Cost: 0.2 GB and about thirty minutes of attention; the 3-bit speed test it was meant to enable was never run.
- Listed as a model. Later inventories list it as 'MLX qwen3_5_mtp 203M' and, on 2026-09-11, as 'one weight file' under models. Nothing runs it.
What happened to it
Downloaded 2026-09-07 → identified as a head 2026-09-07 20:32Z → kept → on disk at models/qwen3.8-27b-mlx-3bit-mtp and in the HF cache stub models--lukaskremla--Qwen3.8-27B-MTP-3bit-MLX (read 2026-10-23). Re-downloadable from HuggingFace.