Qwen3-Next-80B-A3B-Instruct REAP · file← Qwen3-Next-80B-A3B-Instruct REAP
safetensors source
The REAP authors' full-size original. Never run; downloaded twice to make the MLX copies. SamsungSAILMontreal.
- Size
- 112.4 GB
- Status
- deleted
- 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-Next-80B-A3B-Instruct-REAP safetensors, 112.4 GB, 25 shards |
| Format | bf16/fp16 safetensors (dtype never read) |
| Where from | downloaded twice from SamsungSAILMontreal/...-REAP: 2026-09-08 (deleted after conversion) and 2026-09-10 (deleted 2026-09-11 ~09:13Z) |
| Status | deleted |
| Hugging Face | https://huggingface.co/SamsungSAILMontreal/Qwen3-Next-80B-A3B-Instruct-REAP |
Where it came from
- Repo:
SamsungSAILMontreal/Qwen3-Next-80B-A3B-Instruct-REAP, the REAP authors' own checkpoint ("25 downloads" when found). 112.4 GB, 60B parameters, a REAP-60B pruned from 80B, unquantised, 384 experts, 48 layers; 25 shards, fp16/bf16 (the dtype was never read;RESTORE-deleted-20260911.sh). The assistant noted a 124 GB figure at the end of the first download; the extra is not explained (not on record: the partial-file or cache overhead would have been in the directory listing). - Why it was needed:
mlx_lm.ggufonly exports, so GGUF cannot be converted to MLX; the MLX builds had to come from safetensors. - First download (2026-09-08): queued behind the Flash-Next download by
eval/reap_mlx.sh(hf download SamsungSAILMontreal/Qwen3-Next-80B-A3B-Instruct-REAP --local-dir reap-src,HF_HUB_ENABLE_HF_TRANSFER=1,HF_HUB_DISABLE_XET=1), after the assistant freed disk by deletingmuse-glimmer-30b-metaandornith-1.5-35b-mlx-4bit(43 GB). 70 of 115 GB at 17:48Z; finished before 19:37Z. Asked by Alan at 19:02Z: "So reap mlx is next then". - Second download (2026-09-10): to
models/qwen3-next-reap-safetensors/, a per-file range downloader (a one-off script,dl_reap.py) queued behind the Q4_K_M GGUF at 14:13Z and running by 15:16Z; reconverted at 23:52Z. Alan: "reap has their safe tensors available which is how we did it". Revision: not on record either time.
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) | n/a | n/a | n/a | n/a | conversion source only; no run-log label |
Seats held: none.
Decisions
- Fetched to give REAP "under all three runtimes", and because "a mixed-precision REAP build is the more interesting target" than a plain one.
- Deleted after the first conversion by the script itself (
rm -rf reap-src, "source no longer needed"); the assistant reported "source deleted 112 GB reclaimed". - Fetched again on 2026-09-10 to make a controlled reconversion at mixed36's base settings. Deleted 2026-09-11 09:12:18Z on Alan's "nuke it", after the assistant had confirmed the conversion was sound and the source "redundant".
What went wrong
- A 112 GB download made twice. The first source was deleted by design right after conversion, which is why the second conversion needed it again. The second build used the same source but its recipe differs from the first conversion (overrides), so the second download was the price of the controlled test. The repeated 40 GB-class download is the protocol failure the lab's notes record for this family; the 2026-09-11 RESTORE note flags the downloader caveat (single stream ~80 Mbps).
- Disk guard needed. The first pipeline would have failed halfway; the assistant added a free-disk check (
NEED=$(( 112 - HAVE + 52 ))) toreap_mlx.shbefore starting. - Standing checks. Partial runs, stale build stamps, missing flags: n/a (never served).
What happened to it
Downloaded 2026-09-08 → converted (two builds) → deleted by script 2026-09-08 ~20:06Z → downloaded again 2026-09-10 → converted 23:52Z → deleted 2026-09-11 ~09:13Z (93 to 205 GB free). RESTORE-deleted-20260911.sh entry exists with the repo, the downloader advice and the reconversion settings (q_bits=4, q_group_size=64, q_mode=affine, routers land at 8-bit). Re-downloadable from the same repo, ~112 GB.