Qwen3-Coder-Next · file← Qwen3-Coder-Next
Qwen3-Coder-Next IQ4_XS
Unsloth's IQ4_XS, the same format as its Instruct sibling so the scores compare. Deleted on 8 September. Unsloth GGUF.
- Size
- 42.7 GB
- Status
- Deleted 8 Sep 2026
- Scored runs
- 3
- Smallest Mac
- 64 GB estimate, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| groundingqwen3-next-80b-coder | 76.8% | 68.7–83.3 | Passed 96/125 | llama.cppextraction | 2026-08-27 |
| nullschemaqwen3-next-80b-coder | 56.9% | 45.4–67.7 | Passed 41/72 | llama.cppextraction | 2026-08-27 |
| screeningqwen3-next-80b-coder | 90.0% | 79.9–95.3 | Passed 54/60 | llama.cppextraction | 2026-08-27 |
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-Coder-Next-IQ4_XS.gguf |
| Format | GGUF IQ4_XS, 42,676,135,968 B |
| Where from | downloaded 2026-08-27 from unsloth/Qwen3-Coder-Next-GGUF, single-stream curl -C - |
| Status | deleted 2026-09-08 14:05Z; preset commented out |
| Hugging Face | https://huggingface.co/unsloth/Qwen3-Coder-Next-GGUF |
Where it came from
- Repo and file:
unsloth/Qwen3-Coder-Next-GGUF/Qwen3-Coder-Next-IQ4_XS.gguf(not theUD-IQ4_XSvariant, which was listed beside it). - Method: a one-off script, a single
curl -L --retry 30 --retry-delay 5 --retry-all-errors -C -ofresolve/mainintomodels/qwen3-coder-next/, chosen because parallel ranges throttle when unauthenticated. Rate 41 Mbps against 150 Mbps the night before, ETA ~132 min; actual 07:12Z to 08:17Z. - Verified: size equality against the HF API and the GGUF magic bytes (
47475546). - Who asked: Alan, "Pull it also".
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
qwen3-next-80b-coder | 2026-08-27 08:26Z to 08:32Z (run-log finish 09:29:15, 09:31:15, 09:32:10 BST) | llama-server b10618 eb25b7263 (run log) | c = 262144, np = 1, load-mode = none; KV type, -fa, --jinja, --reasoning-format not set on the preset or the router line (--models-preset... --models-max 1 --context-shift) | extraction | third arm of the Instruct / Thinking / Coder trio: same quant, ctx and build as the other two |
Other servings with no run-log label:
- the gate probe (template plus one tool call,
reasoning chars: 0), 08:26Z; - DSH goal-loop sessions on the router, from 2026-08-27 ~15:07Z, last request Thursday 2026-08-27 20:07 (time zone as the assistant reported it); 193,443-token last turn at 14.7 t/s;
- resident and idle until at least 2026-08-31 (restored by the assistant after small-model tests, 02:56Z to 02:57Z, 'box restored: coder loaded').
Seats held: none formal. It was not set to load at startup; it was Alan's chosen DSH model for the stock-platform goal run, and it stayed loaded as the router's resident model for days.
A 2026-09-05 KV experiment edited the Instruct preset to cache-type-k/v = q8_0; whether the Coder preset was changed is not on record (the result cuts off after the header line).
Decisions
- Why fetched: see the family doc (Alan's hunch about coder models, then the three-model experiment).
- Why this file: same IQ4_XS quant and ~46.5 GB resident at 262K as the Instruct night, so all numbers compare.
- Why kept: it was in active use as Alan's DSH model for the stock-platform runs and left loaded; no seat decision was recorded.
- Why deleted: disk pressure on 2026-09-08 (Alan: "start putting models somewhere or deleting more aggressively"). The assistant's rationale in the list was "zero results, never once benchmarked", which was false. Alan's response at deletion time: "we can revisit coder models later".
What went wrong
- False "never benchmarked" claim and a silent zero from the pre-delete check. The assistant's 2026-09-08 check counted result files by the directory name and printed 0 for qwen3-coder-next because the files are named
qwen3-next-80b-coder. The deletion ran anyway. The three result files and three run-log rows were intact, so nothing measured was lost. - Idle residency. The model held 41.55 GB for 3 days 6 hours with no requests, blocking other benchmarks that refuse to start under 36 GB reclaimable.
- Router preset used as a benchmark rig. The suites required evicting the resident Instruct model, which was also Alan's live DSH model, for ~15 minutes (announced).
What happened to it
Downloaded 2026-08-27 07:12Z to 08:17Z, registered 08:17:56Z, benched 08:26Z to 08:32Z, used in DSH from ~15:07Z, left resident through at least 2026-08-31, deleted 2026-09-08 14:05:35Z with qwen3-next-80b-a3b-thinking and qwen-agentworld-35b (101 GB), preset commented out ('; --- pruned 2026-09-08: weights deleted ---'). RESTORE-deleted-20260908.sh function coder re-downloads from unsloth/Qwen3-Coder-Next-GGUF (the right publisher, unlike its thinking sibling). Run-log results survive. No ladder or swap measurements exist to restore.