Qwen3-Next-80B-A3B-Instruct · file← Qwen3-Next-80B-A3B-Instruct
Q4_0
A copy in a format we expected to be faster. Never scored. It ran our Hermes agent for twelve minutes. Unsloth.
- Size
- 45.3 GB
- Status
- Deleted 5 Oct 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 | qwen3-next-80b-gguf/Qwen3-Next-80B-A3B-Instruct-Q4_0.gguf |
| Format | GGUF Q4_0, 45,307,878,176 B (Content-Length) |
| Where from | downloaded 2026-09-11 from unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF |
| Status | deleted 2026-10-05 10:06Z |
| Hugging Face | https://huggingface.co/unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF |
Where it came from
- Repo / file:
unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF→Qwen3-Next-80B-A3B-Instruct-Q4_0.gguf. Content-Length 45,307,878,176 B (42.2 GiB; the "42.20 GB" reported at the time is GiB). - Command: a one-off script used six parallel
curl -rrange requests, then a size check and aGGUFmagic check, intomodels/qwen3-next-80b-gguf/. The script's "SIZE OK" line is not on record. Revision and sha256 are not on record. - Date: 2026-09-11, launched 19:23Z. Complete before the 20:35Z bench.
- Who asked / why: Alan, 16:04Z: "one model we haven't revisited is the qwen3 next 80b with the q4_0". At 19:20Z: "also we definitely need to do the 80 at 4_0 again and you never did respond to my comment about always preferring guffs". The hypothesis, written into the script, was the day's byte-alignment finding: "i-quants are the MOST" … "expensive format to unpack, so today's alignment finding predicts a real gain" … "over IQ4_XS's 41.4 tok/s".
How it was served
No run-log label. Never board-scored.
| role | dates (UTC) | binary | flags | why |
|---|---|---|---|---|
| speed bench | 2026-09-11 20:35Z | .local/bin/llama-cli → b10893 c32d1dabe | -c 4096 -ngl 99 -n 192 --temp 0 -st -v, default mmap | the Q4_0 lever |
| runtime bench vs MLX | 2026-09-12 08:48Z | same, b10893 | three workloads | "I definitely want to test the 80b guff vs mlx" |
| context ladder | 2026-09-12 09:44–09:49Z | llama-server, b10893 | descending 262144 → 8192; load, generate, measure | Alan: "Yes measure" |
| Hermes main + 10 aux roles | 2026-09-12 09:56Z → 10:08Z | .local/bin/llama-server (stock; "this arch is upstream") | --jinja -c 131072 -np 1 -fa on -ngl 999 --alias qwen3-next-80b, f16 KV. Changed to -c 65536 at 10:07Z | Alan: "Not just now, wire 80b into Hermes I’ll go to me desk and play" |
Seat held: Hermes, 2026-09-12 09:56Z → 10:08Z, replaced by ornith-1.5-9b. There was no --load-mode none on any of these lines, although the lab's notes call it mandatory for this architecture.
Decisions
- Downloaded to test the alignment lever on "the board's best abstainer". The speed gain over IQ4_XS was not claimed, because the baseline was on another build (finding).
- Seated in Hermes after the ladder. The assistant: "That's the strongest daily-driver case on your board".
- Unseated after 12 minutes. Alan: "let's try and drive with ornith 9b". The argument made for the swap used a toolcall 78.3% that belongs to REAP mixed36, not to this file or checkpoint (family Decision log).
- Deleted 2026-10-05 10:06Z to free disk for the Kolibri Q8_0 → Q4_0 build, with Alan's approval. The question mis-described it as an "IQ4_XS GGUF … later re-pull" of a model "deleted 2026-08-14".
- Not published (third-party quant).
What went wrong
- Unscored file presented with another file's scores. The
swap-to-80b.shheader says "Board position: nullschema 97.2% (top), synthesis 100% (top), tagger 93.3%". The 97.2 and 93.3 are the IQ4_XS. synthesis 100% is REAP mixed36; this checkpoint never ran synthesis. - Script/header drift:
swap-to-80b.shechoes "(-c 131072 -np 1, fa on)" and marks 131072 "<- chosen", but launches-c 65536. The 10:07Zsedchanged only the command line. - Readiness bug in the ladder's first pass.
curl -sexits 0 on a 503 "loading", so every rung reported GEN_FAIL with RSS sampled mid-load. It was fixed by waiting for"status":"ok"and re-run. - Deleted on stale premises (2026-10-05). The assistant said "the runbook recorded that model as deleted on 14 Aug", and that was taken to mean these were stray re-pulls.
What happened to it
Downloaded 2026-09-11 → speed-benched 09-11 and 09-12 → context ladder 09-12 → Hermes seat 09-12 09:56–10:08Z → idle → deleted 2026-10-05 10:06Z. There is no RESTORE entry. swap-to-80b.sh still points at the missing path. Re-downloadable from unsloth/Qwen3-Next-80B-A3B-Instruct-GGUF.