Qwen3-Next-80B-A3B-Thinking · file← Qwen3-Next-80B-A3B-Thinking
Qwen3-Next-80B-A3B-Thinking IQ4_XS
Unsloth's GGUF, the only copy we held. Tested overnight on 27 August; deleted on 8 September. Unsloth GGUF.
- Size
- 42.6 GB
- Status
- Deleted 8 Sep 2026
- Scored runs
- 4
- Smallest Mac
- 64 GB estimate, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| groundingqwen3-next-80b-thinking | 65.6% | 56.9–73.4 | Passed 82/125 | llama.cppextraction | 2026-08-27 |
| groundingqwen3-next-80b-thinking | 100.0% | 20.7–100.0 | Passed 1/1 | llama.cppstudy-default | 2026-08-27 |
| nullschemaqwen3-next-80b-thinking | 58.3% | 46.8–69.0 | Passed 42/72 | llama.cppextraction | 2026-08-27 |
| screeningqwen3-next-80b-thinking | 91.7% | 81.9–96.4 | Passed 55/60 | 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-Next-80B-A3B-Thinking-IQ4_XS.gguf |
| Format | GGUF IQ4_XS, 42,644,480,160 B |
| Where from | downloaded 2026-08-27 from unsloth/Qwen3-Next-80B-A3B-Thinking-GGUF, single-stream curl -C - |
| Status | deleted 2026-09-08 14:05Z; preset commented out, not removed |
| Hugging Face | https://huggingface.co/unsloth/Qwen3-Next-80B-A3B-Thinking-GGUF |
Where it came from
- Repo and file:
unsloth/Qwen3-Next-80B-A3B-Thinking-GGUF/Qwen3-Next-80B-A3B-Thinking-IQ4_XS.gguf. The assistant checked both unsloth andbartowski/Qwen_Qwen3-Next-80B-A3B-Thinking-GGUFfor an IQ4_XS and chose unsloth to match the Instruct filename and publisher. - Size: 42,644,480,160 bytes (the
SIZEin thecurlcommand, the post-download check expects exactly this). sha256 and HF revision: not on record. - Method: a single detached
curl -L --retry 30 -C -againstresolve/mainintomodels/qwen3-next-80b-a3b-thinking/, after three failed attempts:hf download(stalled at 5.9 MB in 90 s), six range streams on one signed URL (8 Mbps aggregate), six streams on per-stream signed URLs (4 Mbps). Unauthenticated HF throttled parallel ranges below a single stream that night; the stream then ran 98 to 150 Mbps. - Dates: started 2026-08-27 00:23:59Z, complete 01:11:17Z. Asked by Alan ("download it, test it, test the swap times, the memory profile, all of it").
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-thinking | 2026-08-27 02:03Z to 02:57Z (run-log finish 03:03:51 to 03:57:43 BST) | llama-server b10618 eb25b7263 (run log 0.2.0-dev (build 10618, commit eb25b7263)) | c = 262144, np = 1, load-mode = none from the preset; KV type, -fa, --jinja and --reasoning-format are not set on the router line or preset (router: --models-preset fleet-models.ini --models-max 1 --context-shift, launchd plist) | extraction (4 rows) and study-default (1 row) | the single label for the Thinking checkpoint: matched to the Instruct baseline's extraction request shape, plus a 10,000-token arm to give reasoning room |
Other servings with no run-log label (all this checkpoint, none scored on the board):
- swap timing,
eval/swaptime.py, router, 2026-08-27 01:13Z to 01:15Z; - temperature probe,
benchmarks/temp_probe.py, 50 cases at t=0 and t=0.6, 4096 tokens, ~62 min, 02:58Z to 03:59Z; - context ladder,
eval/ctx_ladder.py: six directllama-serverinstances with the router stopped, 04:00Z to 04:02Z; - bloglanding generation through DSH, 04:02Z to 04:12Z.
Seats held: none. The preset [qwen3-next-80b-thinking] did not load at startup; it was reachable on demand from 2026-08-27 to 2026-09-08 (fleet-models.ini).
Router launch has no --jinja on its command line and the preset does not add it; whether build 10618 defaults it on is not on record (would be in the server log, models/router.log).
Decisions
- Why it was fetched: to test whether reasoning improves extraction (family doc, Why it was tested).
- Why it was kept, then not: kept after the verdict because "Keeping it costs nothing" in disk terms (40 GB against 257 GB free, no load at startup, 13 s swap). Deleted in the 2026-09-08 sweep; Alan then said he had expected to keep it as a general reasoning model, and the assistant took the blame for listing it as "rejected".
- Side effects of the night's work (reported to Alan at 06:30Z):
.dsh/profiles/web/cordis.patch.ymlgained the two Qwen3-Next entries;settings.yamlchanged the default DSH model to Qwen3-Next (both with.bak-20260827); the router was taken down entirely for the ladder.
What went wrong
- Wrong model resident, wrong diagnosis. Mid-run the assistant saw 6.3 t/s and falling RSS and blamed page-cache eviction from the concurrent download; the router had swapped in
qwen3.8-27b(a lab note, "Two process failures"). Cost: a wrong interim reading; fixed by checking/v1/modelsbefore interpreting numbers. - Early cost figure wrong twice. First about 13 minutes of generation (the output ceiling mistaken for the running cost), then 9.4x slower extrapolated from ten easy cases, corrected to 25x on the full suite.
- Empty-content probe. The first swap probe at
max_tokens=8returned empty content because all eight tokens went to reasoning. - Standing checks. Partial runs: only
study-default(above). Stale build stamp: none; the run log's b10618 matches the router restarted on 2026-08-26. Missing flags:--jinja/--reasoning-formatnot on the router line (above).
What happened to it
Downloaded 2026-08-27 00:24Z to 01:11Z, benched 02:03Z to 02:57Z (suites), 03:00Z to 04:12Z (probe, ladder, generation), registered on the router, deleted 2026-09-08 14:05:35Z in the sweep that also removed qwen3-coder-next and qwen-agentworld-35b (101 GB, 76 to 176 GB free). The preset was commented out, not removed (; --- pruned 2026-09-08: weights deleted ---). Entry in RESTORE-deleted-20260908.sh function thinking (hf download bartowski/Qwen_Qwen3-Next-80B-A3B-Thinking-GGUF --include "*IQ4_XS*"), but the file held was unsloth's; the re-download source in the script is a different publisher (the filename would carry a Qwen_ prefix). Re-downloadable from unsloth/Qwen3-Next-80B-A3B-Thinking-GGUF. Run-log results survive.