Gemma 4 E2B · file← Gemma 4 E2B
vision projector GGUF
The vision part that came with it. Never loaded, so its vision is untested here. Gemma-4-E2B-it-mmproj.gguf.
- Size
- 1.0 GB
- Status
- On disk
- Scored runs
- 0
This file never completed a scored run. Its record below says why.
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 | gemma-4-E2B-it-mmproj.gguf |
| Format | GGUF vision projector, 986,833,664 B |
| Where from | downloaded with the above, same get.sh |
| Status | on disk; never loaded on record |
| Hugging Face | as above |
Where it came from
- Repo / file:
google/gemma-4-E2B-it-qat-q4_0-gguf/gemma-4-E2B-it-mmproj.gguf, listed at 0.99 GB. - Command: the same
models/gemma4-e2b/get.shloop as the Q4_0 language model; see Gemma 4 E2B — Google QAT Q4_0 GGUF (gemma-4-E2B_q4_0-it.gguf) for the script. - Size: 986,833,664 B (stat, 2026-10-23). sha256 and revision not on record.
- Date: 2026-08-23 ~10:48Z (mtime "23 Aug 11:48" local).
- What it is: the vision tower/projector paired with the language GGUF for llama.cpp's multimodal path (
--mmproj). It is not a quant of the language model and carries no text-generation weights. Its tensor types are not on record.
How it was served
No run-log label. No server on record was ever started with --mmproj pointing at this file. eval/bench_fleet.py was patched on 2026-08-23 specifically so it would not be benched as a model ("four projectors in this fleet are named... gemma-4-E2B-it-mmproj.gguf -- the marker is in the middle, so a startswith filter lets them through and benches a vision tower").
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
| — (none) | — | — | — | — | never served |
Seats held: none.
Decisions
- Downloaded alongside the language model because
get.shfetched both files of the repo. No separate reason is on record (reconstructed: Alan noted on 2026-08-23 that the family is multimodal — "I already know it's multimodal so I get vision too", said of the 12B). - Kept (0.99 GB, not in any sweep).
What went wrong
- Nearly benched as a model by the fleet speed sweep; fixed by the substring filter above.
What happened to it
Downloaded 2026-08-23 → never loaded → on disk at models/gemma4-e2b/gemma-4-E2B-it-mmproj.gguf (986,833,664 B, checked 2026-10-23). Re-downloadable from google/gemma-4-E2B-it-qat-q4_0-gguf.