Qwen3-Next-80B-A3B-Instruct REAP · file← Qwen3-Next-80B-A3B-Instruct REAP
MLX 8-bit
An 8-bit MLX copy, stopped within a minute as too big to load. Abandoned, never completed.
- Size
- 29.0 GB
- Status
- Deleted 10 Sep 2026
- Scored runs
- 0
This file never completed a scored run. Its record below says why.
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-next-reap-mlx-8bit (partial, 29 GB written) |
| Format | MLX 8-bit, abandoned |
| Where from | made here 2026-09-10, killed after ~40 s |
| Status | deleted 2026-09-10 23:56Z |
| Hugging Face | not on HF |
How it was made
- Input: the Samsung safetensors (Qwen3-Next-80B-A3B-Instruct REAP safetensors source (SamsungSAILMontreal), second copy).
- Command: a scratch file, the 4-bit script with
q_bits=8substituted bysed(convert(SRC, mlx_path=DST, quantize=True, q_bits=8, q_group_size=64, q_mode="affine"),DST=models/qwen3-next-reap-mlx-8bit), run asnohup mlx-main-venv/bin/python convert_reap8.py. - Date: started 2026-09-10 23:56:05Z, killed 23:56:47Z. About 29 GB had been written (
partial output: 29G) of an expected ~60 GB. - Who asked: Alan, 23:55:59Z: "In fact let’s try a q8". He withdrew it 37 seconds later: "In fact that’s too big a file to load", then "It has to be q4".
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; no run-log label) | n/a | n/a | n/a | n/a | the conversion was killed before completion |
Seats held: none.
Decisions
Built on a passing remark of Alan's and abandoned on his next message once the file-size arithmetic was on the table; recorded because the 2026-09-10 sweep logic ("Q8 is the ceiling, not BF16", from a lab note on quantisation) was derived on smaller models and does not carry to a 60B MoE on this box. Not published.
What went wrong
- Disk churn: ~29 GB written and removed in 42 seconds (
rm -rfof the partial directory); disk free returned to 134 GB. - Standing checks: partial runs, stale stamps, missing flags and published mismatches: n/a (nothing served or scored).
What happened to it
Started and killed 2026-09-10 23:56Z; partial output deleted at 23:56:47Z. No RESTORE entry and none needed: the recipe is the 4-bit command with q_bits=8. The file never held a usable weights set; it is recorded only because the convention asks for builds made and deleted without a run.