BF16 safetensors

Qwen's full-size original. Used only to make the three MLX files. Qwen, HF cache.

Size: 18.0 GB
Status: Deleted 12 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	HF cache `models--Qwen--Qwen3.5-9B`
Format	BF16 safetensors, 4 shards, 18.0 GB
Where from	downloaded 2026-09-11 by `mlx_lm convert`
Status	deleted 2026-09-12 07:45Z
Hugging Face	https://huggingface.co/Qwen/Qwen3.5-9B

Where it came from

Repo: Qwen/Qwen3.5-9B, revision c202236235762e1c871ad0ccb60c8ee5ba337b9a; 4 safetensors shards, 18.0 GB total. Held in .cache/huggingface/hub/models--Qwen--Qwen3.5-9B (18 GB when deleted), not under models.
Fetched by: mlx_lm convert --hf-path Qwen/Qwen3.5-9B... inside a one-off script (queued 2026-09-11 14:19Z), whose conversion step fetches the repo. The fetch time is not otherwise logged; the weights were present by 14:44Z (the conversion "Quantized model with 4.501 bits per weight"). (reconstructed)
Snapshot defect: the cache lacked .gitattributes, LICENSE, README.md, so mlx_lm's save step (snapshot_download(local_files_only=True)) raised IncompleteSnapshotError after quantising. Fixed on 2026-09-11 15:26Z by hf_hub_download of those three files. sha256 not on record.
Who asked / why: the assistant needed a source for the MLX 4-bit comparison against Q4_0 (the three-way runtime test), then for the three MLX formats.

How it was served

label	dates (UTC)	binary and version	ctx / -np / KV / -fa / `--jinja` / `--reasoning-format`	harness profile	why this label exists
none	n/a	n/a	never loaded as BF16; read only by `mlx_lm convert`	n/a	conversion source for Qwen3.5-9B MLX affine 4-bit, group 64 (converted here), Qwen3.5-9B MLX affine 4-bit, group 32 (converted here), Qwen3.5-9B MLX mxfp4 (converted here)

Seats held: none. The 17.9 GB BF16 GGUF in the same unsloth repo (listed) was never fetched, which is why the quantlab GGUF files were re-rounded from the Q8_0.

Decisions

Kept in the HF cache after conversion so the three MLX formats could be built from the same snapshot. Its deletion was bundled into the "kill" command of 2026-09-12.

What went wrong

Three failed conversions (2026-09-11 14:44Z and 15:25Z) each quantised successfully and then failed at save; the cause was misdiagnosed as disk pressure on the granite model first.
The first _mlxtest/qwen3.5-9b-mlx-4bit output never existed for the same reason, though the assistant first said it "actually succeeded".

What happened to it

Fetched 2026-09-11 (by mlx_lm convert) -> three MLX conversions -> deleted 2026-09-12 07:45:56Z. No RESTORE-deleted-*.sh entry exists for it. Re-downloadable from Qwen/Qwen3.5-9B (revision may have moved on).