Qwen3-Coder-30B-A3B-Instruct · file← Qwen3-Coder-30B-A3B-Instruct
Qwen3-Coder-30B-A3B-Instruct Q4_K_M
The one we kept: unsloth's Q4_K_M, still on disk and ready to load. Unsloth.
- Size
- 18.6 GB
- Status
- On disk
- Scored runs
- 6
- Smallest Mac
- 32 GB estimate, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| groundingqwen3-coder-30b | 77.6% | 69.5–84.0 | Passed 97/125 | llama.cppextraction | 2026-08-24 |
| groundingqwen3-coder-30b-b10618 | 76.8% | 68.7–83.3 | Passed 96/125 | llama.cppextraction | 2026-08-26 |
| nullschemaqwen3-coder-30b-b10618 | 37.5% | 27.2–49.0 | Passed 27/72 | llama.cppextraction | 2026-08-26 |
| screeningqwen3-coder-30b-b10618 | 91.7% | 81.9–96.4 | Passed 55/60 | llama.cppextraction | 2026-08-26 |
| shapegenqwen3-coder-30b | 0.0% | 0.0–24.3 | Passed 0/12 | llama.cppextraction | 2026-09-07 |
| sitebuildqwen3-coder-30b | 0.0% | 0.0–79.3 | Passed 0/1 | llama.cppagentic-fast | 2026-09-07 |
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-30B-A3B-Instruct-Q4_K_M.gguf |
| Format | Q4_K_M GGUF, 18,556,689,568 B |
| Where from | downloaded 2026-08-12 (before our session records) via get-coder.sh |
| Status | on disk; [qwen3-coder-30b] router preset (c = 81920) |
| Hugging Face | https://huggingface.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF |
Where it came from
- Repo / file:
unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf. - Command:
$PG "$B/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf" Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf 6, where$PGis a 6-stream parallel range-fetch helper from an earlier working session andB=https://huggingface.co/unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF/resolve/main. - Size: 18,556,689,568 B (stat, 2026-10-23). sha256: not on record (would have been in an earlier working session or the
pget.shoutput). - Date: 2026-08-12. File mtime 17:37:40 and
.dl_complete17:37 (stat 2026-10-23). This predates our session records. - Revision: not on record.
- Who asked / why: the reason in the script header is "pairs with Agents-A1 (21.2GB) under the 51.8GB Metal working set so planner+coder can both stay resident". Who asked is not on record (before our session records).
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
qwen3-coder-30b (grounding) | 2026-08-24 16:45Z – 16:54Z | llama.cpp/build/bin/llama-server b10369 (6e62ba538) | -c 32768 -np 4 -cb -ngl 99; KV default; -fa not set; --jinja; no reasoning flag | extraction | first gauntlet pass after Alan raised it |
qwen3-coder-30b (sitebuild, shapegen) | 2026-09-07 22:35Z – 22:51Z | router llama-server --models-preset fleet-models.ini --models-max 1, b10618 per run-log stamp (runtime_source "router") | preset: c = 81920; child launched with --context-shift --jinja --ctx-size 81920 --n-gpu-layers 99 (child command line seen 2026-09-08); -np, KV, -fa not set | sitebuild: agentic-fast (thinking on, 4096 tokens, budget 512); shapegen: extraction | first coding suites; it was the router's resident coder |
qwen3-coder-30b-b10618 | 2026-08-26 ~06:5xZ – 07:03:47Z | b10618 (eb25b7263) | -c 32768 -np 4 -cb -ngl 99 --jinja | extraction | fleet re-run on the new build; label suffix keeps builds apart |
(not in run log) Toolery qwen3-coder-30b, -t07, -b10618 | 2026-08-24 16:45Z; 2026-08-26 | b10369 / b10618 | as above | Toolery raw, all tiers; adversarial six at temperature 0.7 × 5 | gauntlet stages |
| (not in run log) frontend bench, coresidency, plan/execute | 2026-08-19 (before our session records) | b10369 | preset at 81920 | bench_frontend.py via dsh; coresidency.py | two lab notes |
| (not in run log) Hermes/opencode/dsh use | 2026-08-19 → 2026-09 | router | preset | — | cortex/qwen3-coder-30b: 15 messages in opencode's history |
Seats: never a fleetview seat. It has been a router preset in fleet-models.ini throughout: c = 81920 from 2026-08-19 (ini backups from 2026-08-19 through 2026-09-08, and live). The 2026-09-12 inventory called it an "active fleet seat" because it was a preset.
Timezone note: run-log ts 2026-08-24T17:54:17 BST = 16:54Z; 2026-09-07T23:38:29 / 23:51:55 BST = 22:38Z / 22:51Z, matching the commands.
Decisions
- Downloaded as the coder half of a co-resident pair with Agents-A1 (
get-coder.sh). - Context capped at 81920 on 2026-08-19 for headroom and
--cache-ram(ini comment). - Kept as a preset but given no seat. It was ruled out for untrusted content and not preferred for coding over qwen3-next-80b. Coder roles were parked by Alan on 2026-09-08.
- No Q4_0 made, though
fleet_auditasked for one. The audit was later rescoped for MoE. - Published on the site as
qwen3-coder-30b(status candidate).
What went wrong
- Superseded shapegen/1 and sitebuild smoke (2026-09-07). shapegen/1 was graded under a grader later found to miscount primitives as shape nesting; sitebuild's first grader crashed on a model-emitted
interfaceslist of strings. Both were fixed before the run-log rows above; the earlier files are kept as.superseded. - Misdescribed architecture in an eval config: "Qwen3 Coder dense", active_params "30B".
- Stale speed in docs: 52.5 t/s, corrected on 2026-08-23 to 64.4. It still appears in a lab note's model table, and in a table inside a gold-240 dataset case.
- Coin-pusher claim reversed on 2026-08-27 ("working Coulomb-friction physics"). The 2026-08-19 play-test record says the pusher never touches a coin (a lab note).
- Router residue (2026-09-08 15:38Z). The router reported it "loaded" while its child had 0.0 GB RSS, evicted to swap; the router was restarted.
- Standing checks: no partial runs. The run-log build stamps agree with the evidence (b10369 for 08-24, b10618 for 08-26). The router rows rely on the router's own stamp.
--jinjapresent on every launch on record.
What happened to it
Downloaded 2026-08-12 (before our session records) → router preset (81920) 2026-08-19 → gauntlet 2026-08-24 (b10369) → re-run 2026-08-26 (b10618) → coding suites 2026-09-07 (router) → protected in the 2026-09-12 Q4_K_M audit → survived the 2026-10-11 sweeps → on disk at models/qwen3-coder-30b/Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf (18,556,689,568 B, 2026-10-23), still a live router preset. No RESTORE entry (never deleted). Re-downloadable from unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF.