Gemma 4 E4B Hybrid GGUF — Resilient AI Challenge 2026
Team: Visca Barca
Category: Image-to-Text
Base model: google/gemma-4-E4B-it
Challenge: UNESCO Resilient AI Challenge 2026
Results
| Metric | Baseline (unquantized) | This Model |
|---|---|---|
| VQA Accuracy | 46% | 54% |
| Model Size | 15.0 GB | 5.88 GB |
| Size Reduction | — | 61% |
| Accuracy Retention | — | 117% of baseline |
| Avg TTFT | 1.45s | 3.91s |
| CO₂/sample | 0.028g | 0.117g |
Evaluated on 50 stratified VQAv2 samples (15 yes/no, 15 counting, 20 open-ended) on NVIDIA RTX 5070 Ti 16GB.
Files
| File | Description |
|---|---|
gemma4-E4B-hybrid2.gguf | Main model weights — hybrid Q8/Q4_K_M quantization |
gemma4-E4B-mmproj.gguf | Vision projector — multimodal image encoder |
llama_config.yaml | llama-server inference configuration |
Inference
llama-server -hf Defnotirf/gemma-4-E4B-hybrid-GGUF --n-gpu-layers 99
Or with explicit paths:
llama-server \
-m gemma4-E4B-hybrid2.gguf \
--mmproj gemma4-E4B-mmproj.gguf \
--n-gpu-layers 99 \
--port 8080
License
Released under the same license as the original model:
Gemma Terms of Use