MedGemma 1.5 with Native Ollama Thinking

This model is derived from the official Ollama medgemma1.5:latest model and adds native support for Ollama’s thinking interface.

Changes

Compared with the official model:

• Added native reasoning separation through message.thinking.
• Updated the chat template to recognize MedGemma’s internal reasoning delimiters:
  • Opening: <unused94> thought
  • Closing: <unused95>
• Set temperature to 0 for stable delimiter generation and deterministic decoding.
• Added a modification notice to the HAI-DEF license.
• Preserved the original vision capability.

The following components are unchanged:

• Model weights
• Tokenizer
• Architecture
• Q4_K_M quantization

Usage

ollama run jordimurgo/medgemma1.5-thinking:q4_K_M --think

Thinking is exposed separately from the final response. To hide it in the CLI:

ollama run jordimurgo/medgemma1.5-thinking:q4_K_M --hidethinking

API Example

curl http://localhost:11434/api/chat -d '{
  "model": "jordimurgo/medgemma1.5-thinking:q4_K_M",
  "messages": [
    {
      "role": "user",
      "content": "Explain the possible findings in this medical image."
    }
  ],
  "think": true,
  "stream": false
}'

The response uses separate fields:

{
  "message": {
    "role": "assistant",
    "thinking": "Internal reasoning trace...",
    "content": "Final answer..."
  }
}

Model Details

• Base model: medgemma1.5:latest
• Architecture: Gemma 3
• Parameters: 4.3B
• Quantization: Q4_K_M
• Size: approximately 3.3 GB
• Capabilities: text generation, vision and thinking

Important Notice

MedGemma is intended as a foundation model for healthcare and life-science development. It has not been validated for direct clinical use.

Its outputs must not be used directly for diagnosis, treatment, patient management or other clinical decisions without appropriate validation and independent professional review.

MedGemma has not been evaluated or optimized for multi-turn applications.

License

This derivative remains subject to the Health AI Developer Foundations Terms of Use:

https://developers.google.com/health-ai-developer-foundations/terms

HAI-DEF is provided under and subject to the Health AI Developer Foundations Terms of Use. Only the Ollama chat template and decoding temperature were modified; the model weights and tokenizer remain unchanged.