Qwen3.8-Flash-Next · file← Qwen3.8-Flash-Next
AD-3.84bpw IQ4_XS M64
The build made for 64 GB Macs. It fitted in memory and ran well, but lost to the model it was up against. Deleted. AtomicChat GGUF.
- Size
- 84.9 GB
- Status
- Deleted 7 Sep 2026
- Scored runs
- 3
- Smallest Mac
- Over 128 GB estimate, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| groundingqwen38-flash-next | 79.2% | 71.3–85.4 | Passed 99/125 | llama.cppextraction | 2026-09-05 |
| nullschemaqwen38-flash-next | 50.0% | 38.7–61.3 | Passed 36/72 | llama.cppextraction | 2026-09-05 |
| screeningqwen38-flash-next | 96.7% | 88.6–99.1 | Passed 58/60 | llama.cppextraction | 2026-09-05 |
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 | Qwen3.8-Flash-Next-AD-3.84bpw-IQ4_XS-M64-000{01..28}-of-00028.gguf (28 shards) |
| Format | AD 3.84bpw, IQ4_XS-based mix, M64 layout; 84,930,924,160 B (HF listing) |
| Where from | downloaded 2026-09-05 |
| Status | deleted 2026-09-07 17:36Z; no RESTORE entry |
| Hugging Face | https://huggingface.co/AtomicChat/Qwen3.8-Flash-Next-GGUF/tree/main/Qwen3.8-Flash-Next-AD-3.84bpw-IQ4_XS-M64 |
Where it came from
- Repo / path:
AtomicChat/Qwen3.8-Flash-Next-GGUF, folderQwen3.8-Flash-Next-AD-3.84bpw-IQ4_XS-M64/, 28 shardsQwen3.8-Flash-Next-AD-3.84bpw-IQ4_XS-M64-000NN-of-00028.gguf. Shard 2 is 35.76 GB (the isolated n-gram table); shard 1 0.65 GB; the rest ~1.7 GB each. - Local path:
models/qwen38-flash-next-m64/. - Size: 84,930,924,160 B (HF listing). Measured locally: "local total: 79.10 GB remote total: 79.10 GB mismatched/missing: 0". sha256 not on record. Revision not pinned.
- How: a one-off script, a per-shard loop that skips a shard only when its local size equals the remote
content-length, started 2026-09-05 17:42Z. Finished 28/28 by 19:28Z. Rate ~80 Mbps. - Who asked: Alan, "can you look into AtomicChat currently recommends the 3.84 bpw M64 build for 64GB,", then "yeah why not".
- No mmproj was fetched in this download (the script names only the 28 shards).
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
qwen38-flash-next | 2026-09-05 19:29Z load; suites 21:49Z–22:00Z | llama.cpp-next/build/bin/llama-server, 0.4.0-dev (build 30, commit 6a1a922) | --ctx-size 16384 --n-gpu-layers 99 --jinja -np 1; KV type default (f16, not set); -fa not set; no --reasoning-format; --alias qwen38-flash-next | extraction (thinking off via chat_template_kwargs, verified at 21:49Z) | the only label for this build: grounding, screening, nullschema against qwen3-next-80b |
The launch set -np 1 ("np 1 mandatory per the PR"). Harness: harness/runner.py --suite {grounding,screening,nullschema} --model-id qwen38-flash-next --profile extraction --no-resume, then the server was killed and qwen3-next-80b reloaded on the router. The router's qwen3-next-80b was unloaded to make room.
Seats held: none.
Decisions
- Downloaded to test whether a 125B/6B-active model beats qwen3-next-80b (80B/3B) on the board's own suites.
- Not seated. "Results in, and Qwen3-Next keeps its seat" — nullschema 50.0 vs 97.2, grounding 79.2 vs 88.8, screening 96.7 vs 91.7.
- Deleted 2026-09-07 17:36Z as "measured-and-rejected" in a 220.5 GB sweep, on Alan's "Go for it".
- Re-queued and dropped 2026-09-08: queued again at 14:20Z, then replaced by the 4.27bpw build before any bytes landed because its 82.68% top-1 agreement would confound the test.
- Not published to HF (third-party file).
What went wrong
- Empty content at a 300-token budget in the first throughput probe: thinking consumed the budget. The bench used thinking off, so no suite was affected.
- No RESTORE entry. The 2026-09-07 sweep wrote none;
RESTORE-deleted-20260909.shdescribes the 4.27bpw build only (grep, 2026-10-23). - Partial runs: none. Missing flags:
--jinjapresent; no reasoning format, by design (thinking disabled per request).
What happened to it
Downloaded 2026-09-05 17:42Z–19:28Z → loaded and probed 19:29Z → benched 21:49Z–22:00Z → unloaded, router restored 22:00Z → deleted 2026-09-07 17:36Z (79G) → no RESTORE entry → re-downloadable from AtomicChat/Qwen3.8-Flash-Next-GGUF (folder Qwen3.8-Flash-Next-AD-3.84bpw-IQ4_XS-M64). Needs a llama.cpp with qwen4exp (upstream ≥ PR #27742).