---

license: gemma

base_model: google/gemma-4-E4B-it-qat-q4_0-unquantized

tags:

• gguf
• llama.cpp
• gemma4
• multimodal
• medical
• german

---

gemma4-e4b-v13-plainlora-r16 — GGUF

llama.cpp GGUF of Scribion's German medical fact-extraction model — the project champion

(E4B + plain LoRA r16, no QAT, on the v13b dataset).

Files

| file | what | precision |

|------|------|-----------|

| gemma4-e4b-v13-plainlora-r16.Q4_1.gguf | language model | Q4_1 (affine, gs32) — 5.4 GB |

| mmproj-gemma4-e4b-v13-bf16.gguf | vision + audio projector | bf16 — 0.99 GB |

Both built from the QAT base the adapter was finetuned on (bf16 → Q4_1 for the LM; bf16 mmproj

exported directly from that base's vision/audio towers, which the language-only LoRA leaves untouched).

Use Q4_1 or Q4_K_M, never Q4_0 (the symmetric Q4_0 grid degenerates this asymmetric-QAT model).

• Adapter: Mediform/gemma4-e4b-v13-plainlora-r16
• MLX-swift deploy: Mediform/gemma4-e4b-v13-plainlora-r16-mlx-4bit

Q4_1 is the validated non-Apple proxy for the MLX-4bit deploy (reproduces the Mac mlx-swift eval

closely). Beats the stock base on froehlich, ~on par on arztbericht.

Use

# text-only extraction
llama-server -m gemma4-e4b-v13-plainlora-r16.Q4_1.gguf -c 12288 -np 1 -ngl 99
# multimodal (image / audio input) — add the projector
llama-server -m gemma4-e4b-v13-plainlora-r16.Q4_1.gguf \
             --mmproj mmproj-gemma4-e4b-v13-bf16.gguf -c 12288 -ngl 99

Extraction prompts are multi-call (one per section type); temp 0.3.