FreshContextA context lab · Fife
Drawing
02 · Benchmarks
Sheet
2 of 3
Rev
A · 2026-10-01

Gemma 4 12B · file← Gemma 4 12B

mmproj

The add-on that lets it read images. Loaded for a while, never tested, and not used for tagging. Vision projector, Google QAT repo.

Size
0.2 GB
Status
On disk
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.

Filemmproj-gemma-4-12b-it-qat-q4_0.gguf
Formatmultimodal projector (vision/audio towers), 175,115,616 B; tensor type not on record
Where fromdownloaded with the GGUF, same script
Statuson disk; not loaded by the live seat
Hugging Facesame repo

Where it came from

  • Repo / file: google/gemma-4-12B-it-qat-q4_0-gguf / mmproj-gemma-4-12b-it-qat-q4_0.gguf, 0.18 GB in the HF listing.
  • Command: the same get.sh loop as the language model, pget.sh "$B/$f" "$f" 6. File mtime 23 Aug 11:37:49 BST.
  • Size: 175,115,616 B (stat, 2026-10-23). Revision, sha256, tensor type: not on record. The local GGUF header was not read (ls/stat only), so "q4_0" in the name is Google's naming and not evidence of the projector's precision.
  • What it is: the vision (and, per the HF config, audio) tower adapter that llama-server --mmproj loads beside the LM. It is a separate weights file, not a quant of the 12B.
  • Who asked / why: downloaded by default with the LM. Alan already counted vision as part of the 12B's value: "I already know it's multimodal so I get vision too.".

How it was served

It has no run-log label of its own and no suite measures vision. It was loaded alongside Gemma 4 12B QAT Q4_0 (Google GGUF, downloaded) in these servers:

labeldates (UTC)binary and versionctx / -np / KV / -fa / --jinja / --reasoning-formatprofilewhy this label exists
gemma4-12b (synthesis/1, synthesis/2, worldview/2)2026-09-09 16:54–18:41Zllama.cpp b10618--mmproj … --jinja -c 8192 -np 2 -ngl 999extractionresident-fleet launch line included the projector
gemma4-12b-b10893 (synthesis/2)2026-09-12 16:39Zb10893 per /propsreasoning-seat-gemma12b.sh: --mmproj --jinja -c 65536 -np 1 -fa on -ngl 999extractionpairing check with Ornith 9B

Text-only scores under these labels are in the LM's doc. Whether loading the projector changes text output is not on record.

Seats held with the projector:

  • A server 2026-09-09 → 09-12, including the Hermes vision aux role from 2026-09-10;
  • vision-only seat 2026-09-12 09:55–09:57Z;
  • reasoning seat 2026-09-12 10:10Z;
  • A server at q4_0 KV from 2026-09-13 20:44Z, "+ vision, mmproj loaded".

The TAG seat (2026-09-20 →) does not load it (plist args, 2026-10-23).

Decisions

  • Downloaded with the LM by default.
  • Kept loaded for vision while Hermes used local aux roles.
  • Dropped from the permanent tagger seat on 2026-09-20. The launch line omits --mmproj; no reason was stated (reconstructed: tagging is text-only).

What went wrong

  • 2026-09-10: the assistant said "no vision model resident" while --mmproj was on gemma's launch line. Alan's "is vision not in 12b gemma?" corrected it.
  • 2026-09-12 09:57Z: killing gemma left Hermes vision routed to a dead endpoint, "anything in Hermes that sends an image will fail".

What happened to it

Downloaded 2026-08-23 → loaded seats 2026-09-09 → 09-13 → unused since 2026-09-20 → on disk at models/gemma4-12b-dense/mmproj-gemma-4-12b-it-qat-q4_0.gguf (175,115,616 B, checked 2026-10-23). Never deleted, so no RESTORE entry. Re-downloadable from the same HF repo.