🦷 DentalGemma 1.5 4B IT (GGUF)

This repository contains the GGUF quantized formats of DentalGemma 1.5 4B IT. These files are designed for efficient local inference using llama.cpp and compatible frameworks on edge devices, mobile applications (such as React Native with llama.rn), and consumer hardware.

DentalGemma is a fine-tuned version of MedGemma 1.5 4B IT specialized for dental diagnostics and clinical reasoning. It represents a novel domain adaptation, bridging the gap between medical foundation models and localized dental expertise.

Built for the MedGemma Impact Challenge, this GGUF version allows the DentalGemma model to be brought directly into the hands of clinicians and patients via affordable low-resource devices and mobile platforms, entirely offline and privacy-preserving.

🎯 Model Description

DentalGemma is a multimodal vision-language model that combines:

Visual understanding of dental images (clinical photographs and radiographs)
Clinical reasoning for diagnosis, treatment planning, and patient counseling
Structured assessment following evidence-based dental protocols

For full details on the training data, methodology, benchmark performance, and full precision evaluation, please refer to the original model card.

💻 Usage with llama.cpp

The GGUF format enables highly efficient CPU and GPU inference without requiring heavy Python dependencies, PyTorch, or Transformers.

Prerequisites

Install llama.cpp following the instructions in the llama.cpp repository.
Download both the multimodal projector file (mmproj-*.gguf) and the main quantized model file (*.gguf) from the files section.

CLI Inference (Multimodal Vision-Language)

Use the llama-llava-cli tool (or llava-cli depending on your llama.cpp version/build) to process images and text simultaneously:

./llama-llava-cli -m dentalgemma-1.5-4b-it.Q4_K_M.gguf \
    --mmproj mmproj-dentalgemma-1.5-4b-it-f16.gguf \
    -p "<bos><start_of_turn>user\n[img-1]Analyze this dental X-ray for any abnormalities.<end_of_turn>\n<start_of_turn>model\n" \
    --image dental_xray.jpg \
    -c 4096 \
    --temp 0.1

CLI Inference (Text-only Clinical Cases)

Use the standard llama-cli tool for clinical case reasoning and textual interactions:

./llama-cli -m dentalgemma-1.5-4b-it.Q4_K_M.gguf \
    -p "<bos><start_of_turn>system\nYou are an expert dental clinician and radiologist AI assistant.<end_of_turn>\n<start_of_turn>user\nA 35-year-old male presents with severe throbbing pain in the lower right molar region for 3 days. What is your assessment and management plan?<end_of_turn>\n<start_of_turn>model\n" \
    -c 4096 \
    --temp 0.1

Usage in Mobile Applications (React Native)

This GGUF model powers the offline-first dentalgemma-mobile React Native app via llama.rn integration. When deploying to mobile devices:

Copy both the quantized model weights (.gguf) and multimodal projector weights (mmproj.gguf) to the iOS/Android device's local file storage.
Initialize the native completion context pointing to both files.
Note on Memory: Memory usage is highly optimized with quantized models. For example, a 4-bit quant (Q4_K_M) requires approximately ~2.5GB to 3GB RAM, making it feasible for execution on mid-to-high-end smartphones (8GB+ unified memory recommended for multimodal generation).

🏗️ Model Architecture

Image Encoder: SigLIP — processes dental X-ray images into visual embeddings
Language Model: Gemma 3 architecture (4B parameters) — generates clinical text responses
Modality Fusion: Cross-attention mechanism for image-text integration
Supported Context: 8,192 tokens
Quantization: Standard llama.cpp K-quants (e.g., Q4_K_M, Q8_0)

The base model was first fine-tuned in full bfloat16 using LoRA. The learned adapters were then merged into the base weights, and the resulting full-precision model was converted to GGUF format using llama.cpp scripts (convert_image_encoder_to_gguf.py for the projector and convert_hf_to_gguf.py for the LLM).

⚠️ Ethical Considerations & Disclaimer

Not a diagnostic tool: This model is for research and educational purposes only. All AI-generated assessments must be validated by licensed dental professionals.
Quantization Impact: Please be aware that aggressive quantization (e.g., Q3 or below) may degrade the prompt following, diagnostic reasoning, or image interpretation capabilities compared to the full-precision bfloat16 weights. We recommend Q4_K_M or higher for the best balance of speed and accuracy.
Hallucination risk: Like all LLMs, the model may occasionally generate plausible-sounding but clinically incorrect information.

This model is NOT a substitute for professional dental diagnosis. All AI-generated assessments must be validated by licensed dental professionals before any clinical use. The model is provided "as-is" for research purposes only.

📜 License

Model License: Apache 2.0
Base Model: MedGemma 1.5 4B IT (Gemma Terms of Use)

📚 Citation

If you use DentalGemma in your research, please cite:

@misc{dentalgemma2026,
  title={DentalGemma: Fine-tuning MedGemma for Dental Diagnostics},
  author={Syed Naazim Hussain},
  year={2026},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/naazimsnh02/dentalgemma-1.5-4b-it-GGUF}},
  note={MedGemma Impact Challenge submission}
}