Qwen3-Next-80B-A3B-Instruct REAP · file← Qwen3-Next-80B-A3B-Instruct REAP
MLX 4-bit, converted 2026-09-10
A second 4-bit MLX copy, to test the mixed copy's settings. Deleted at the end.
- Size
- 31.6 GB
- Status
- Deleted 11 Sep 2026
- Scored runs
- 7
- Smallest Mac
- 64 GB estimate, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| nullschemareap-60b-mlx-4bit-uniform | 84.7% | 74.7–91.2 | Passed 61/72 | rapid-mlxextraction | 2026-09-11 |
| synthesisreap-60b-mlx-4bit-uniform | 85.0% | 73.9–91.9 | Passed 51/60 | rapid-mlxextraction | 2026-09-11 |
| toolcallreap-60b-4bit-repro | 25.0% | 15.8–37.2 | Passed 15/60 | rapid-mlxextraction | 2026-09-11 |
| toolcallreap-60b-mlx-4bit-uniform | 28.3% | 18.5–40.8 | Passed 17/60 | rapid-mlxextraction-t07 | 2026-09-11 |
| toolcallreap-60b-mlx-4bit-uniform | 25.0% | 15.8–37.2 | Passed 15/60 | rapid-mlxextraction | 2026-09-11 |
| toolscalereap-60b-mlx-4bit-uniform | 95.0% | 83.5–98.6 | Passed 38/40 | rapid-mlxextraction | 2026-09-11 |
| worldviewreap-60b-mlx-4bit-uniform | 47.5% | 32.9–62.5 | Passed 19/40 | rapid-mlxextraction | 2026-09-11 |
Runs marked unenforced were measured on a suite that requires constrained decoding, on a server shown not to apply the schema: mlx_lm accepts it and ignores it, and a few other servers and models break it. Those numbers do not measure what the suite is designed to measure. They are kept and labelled rather than deleted.
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-4bit-uniform, 31.6 GB |
| Format | MLX 4-bit g64 with 96 router overrides at 8-bit, made here |
| Where from | same safetensors, 2026-09-10 23:52Z |
| Status | deleted 2026-09-11 19:36Z |
| Hugging Face | not on HF |
How it was made
- Input: the Samsung safetensors, second download (Qwen3-Next-80B-A3B-Instruct REAP safetensors source (SamsungSAILMontreal),
models/qwen3-next-reap-safetensors). - Command: a scratch file, calling
mlx_lm.convert(SRC, mlx_path=DST, quantize=True, q_bits=4, q_group_size=64, q_mode="affine")with no predicate, matched to mixed36's own declared base settings so that only the override map differs. The 4-bit script text is not on record; its 8-bit sibling was derived from it bysedand shows the call. Interpreter for the 8-bit sibling:mlx-main-venv/bin/python(mlx_lm 0.32.0); that this build used the same venv is reconstructed. - Result: converted in 41 s, 31.6 GB,
bits=4 group_size=64 mode=affine, 96 per-layer overrides, allbits: 8onmlp.gate(48 layers x 2 router gates). The assistant had planned this as "uniform" and corrected itself when it saw the override count. - Date: 2026-09-10 ~23:50Z. Loads in 12.8 s, 7 shards.
- Who asked: Alan, 13:53Z, "nah I want to get the 4bit ream back" and 14:08Z "it should be available"; the assistant could not find a published uniform REAP MLX build (401 on repo guesses) and made one from the safetensors Alan pointed to.
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
reap-60b-mlx-4bit-uniform | 2026-09-11 08:45:52Z toolcall, 08:49:09Z toolscale, 08:50:38Z worldview, 08:52:49Z nullschema, 08:55:19Z synthesis; 09:10:36Z toolcall at temperature 0.7 | rapid-mlx 0.13.4: serve <path> --served-model-name <path> --enable-auto-tool-choice --tool-call-parser qwen3 --reasoning-parser qwen3 | n/a | extraction; the last row extraction-t07 (temp 0.7, top_p 0.95) | the controlled comparison against the deleted mixed36's recorded scores |
reap-60b-4bit-repro | 2026-09-11 09:09:10Z | same | n/a | extraction | reproduce the 25.0 toolcall at temperature 0 on a fresh label with no resume |
Seats held: the new build replaced mixed36 in the Hermes/dsh/labs configs from 2026-09-10 23:58:19Z. The assistant's seating call on 09:03Z was that it "earned a non-agentic seat", at 31.6 GB "a demonstration, not a resident".
Decisions
- Why made: to separate override-map effects from the prune, which had confounded the mixed36 numbers (family doc).
- Reading, as the assistant stated it: "So there is no prune confound" and "Single variable: the per-layer overrides" — after an earlier hour of flagging a possible prune difference. Then reframed in a lab note as a bit-width effect plus a yes/no disposition scored in opposite directions by two suites.
- Why kept: briefly as a verified conversion and, with the Q4_K_M GGUF, "your cleanest runtime-vs-runtime pair".
- Why deleted: Alan, 2026-09-11 19:34:59Z, "Not reap, ditch reap"; the assistant verified 57 result files first and Alan said "That’s fine". Deleted 19:36Z with the Q4_K_M GGUF, 66 GB together.
- Published? Run-log rows, Badly-chosen mixed precision costs abstention; sensitivity-aware mixing protects it (status revised 2026-09-11) and a lab note.
What went wrong
- "Uniform" was not uniform. The assistant's published claim ("uniform beats mixed") used a 4-bit build with 96 router overrides at 8-bit against a build with 542 overrides, 487 of them down to 3-bit. The finding was rewritten on 2026-09-11.
- Wrong comparison pair, caught. The 95.8 nullschema once cited for a "uniform REAP" belongs to the unpruned mlx-community model (
variant-qwen3-next-80b-mlx-4bit); this build's measured 84.7 replaced it in a lab note, which still calls the 95.8 an "unverifiable" differently-built REAP. - Fleet killed by a side test. At 19:35Z a Squish script stopped the whole serving fleet, Hermes models included, before the assistant stopped it. Not specific to this build.
- Standing checks. Partial runs: none (rows match; 15/60-style counts are scores, not coverage). Stale build stamp: rapid-mlx 0.13.4 stamped. Missing flags: none for the board (parsers on).
What happened to it
Converted 2026-09-10 ~23:50Z → repointed into Hermes/dsh/labs 23:58Z → benched 2026-09-11 08:43Z to 10:10Z → deleted 2026-09-11 19:36Z. No RESTORE-deleted-*.sh entry for the build itself; RESTORE-deleted-20260911.sh (written 09:13Z for the source) records the reconversion recipe: q_bits=4, q_group_size=64, q_mode=affine, "routers land at 8 bits automatically: 96 overrides on mlp.gate", and says this build is self-contained. Re-creatable from the safetensors in about a minute plus a 112 GB download.