Qwen3.5-9B · file← Qwen3.5-9B
Q8_0 + MTP head
The Q8_0 with an extra head for drafting tokens ahead. Speed-checked once; the drafting test never ran. Unsloth MTP GGUF.
- Size
- 9.8 GB
- Status
- Deleted 7 Sep 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 | qwen3.5-9b-mtp/Qwen3.5-9B-Q8_0.gguf |
| Format | Q8_0 + nextn MTP head, 9332 MiB |
| Where from | downloaded 2026-08-23 from unsloth/Qwen3.5-9B-MTP-GGUF |
| Status | deleted 2026-09-07 17:36Z |
| Hugging Face | https://huggingface.co/unsloth/Qwen3.5-9B-MTP-GGUF |
Where it came from
- Repo / file:
unsloth/Qwen3.5-9B-MTP-GGUF/Qwen3.5-9B-Q8_0.gguf(9.79 GB on HF's listing). Fetched bymodels/qwen3.5-9b-mtp/get.shwith 6-rangepget.sh;size 9332 MiB in 6 parts. Done 10:57Z. - What it is: the same weights as the plain build plus an embedded MTP (
nextn) draft head; the script comment says the plain and MTP builds carry identical weights, "so this is a matched-pair speculation A/B" rather than a model comparison, and warns that llama.cpp "silently ignores the head" withoutspec-type = draft-mtp. Verified by reading GGUF tensors: 4nextntensors in the MTP Q8_0 against 0 in the plain Q8_0. Parameter count 9,197,093,888 against 8,953,803,264. - Size: llama-bench tensor
model_size9,775,091,712 B; the file's byte size is not on record. sha256 not on record. - Who asked: the assistant offered it as a question; Alan answered "Yes, both quants with MTP".
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
| (llama-bench, no label) | 2026-08-23 13:0x | llama-bench b10369 (6e62ba538) | -p 512 -n 128 -r 5, MTP not engaged (no spec-type) | n/a | check that the dormant head costs nothing: pp512 411.88 +/- 1.43, tg128 31.59 +/- 0.01 |
Seats held: none. No MTP speculation run on this file is on record: the only draft-mtp commands on record are for other models.
Decisions
- Downloaded on 2026-08-23 for a speculation A/B.
- Pre-registered expectation (2026-08-23 12:36Z): because Q4 and Q8 decoded at the same speed, the 9B "isn't bandwidth-bound at all", so speculation should win on it, and the matched MTP pair was "the cleanest speculation A/B available anywhere". The test was "a cheap experiment when the box frees up" and the assistant was "happy to leave it queued"; it was never run. That reasoning was later undercut, since the equality came from K-quant unpacking cost, not an overhead floor.
- Deleted 2026-09-07 17:36Z in Alan's disk sweep ("Go for it"). The assistant had listed the plain and MTP directories as near-identical duplicates; the delete script filed the directory under "superseded / measured-and-rejected", which misdescribes it: nothing had been measured with the head.
What went wrong
- Mislabelled deletion reason, as above: "measured-and-rejected" but never measured.
- Standing checks: no suite ran, so partial-run, build-stamp and flag checks do not apply.
What happened to it
Downloaded 2026-08-23 -> llama-bench only -> deleted 2026-09-07 17:36Z. No RESTORE-deleted-*.sh entry exists for it. Re-downloadable from unsloth/Qwen3.5-9B-MTP-GGUF.