EpistemeAI/Reason-Medical-20b-4bit

TEXT GENERATIONConcurrency Cost:1Model Size:20BQuant:FP8Ctx Length:32kPublished:Jun 18, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

EpistemeAI/Reason-Medical-20b-4bit is a 20 billion parameter decoder-only Transformer causal language model, fine-tuned from openai/gpt-oss-20b, specifically designed for advanced medical reasoning. It was trained on 370,000 high-quality medical question-and-answer examples using Chain-of-Thought reasoning and Unsloth optimization. This model excels in biomedical question answering, medical exam-style reasoning, and clinical knowledge evaluation for research and development purposes.

Loading preview...

EpistemeAI/Reason-Medical-20b-4bit: Medical Reasoning LLM

This model is a specialized 20 billion parameter medical reasoning language model, fine-tuned by EpistemeAI from openai/gpt-oss-20b. It leverages a decoder-only Transformer architecture and was optimized using the Unsloth method for efficient fine-tuning.

Key Capabilities & Training

  • Advanced Medical Reasoning: Designed for professional medicine, medical genetics, college biology/medicine, and clinical knowledge.
  • Chain-of-Thought (CoT) Reasoning: Fine-tuned on 370,000 high-quality medical question-and-answer examples to improve step-by-step problem-solving.
  • Base Model: Built upon openai/gpt-oss-20b, a sparse Mixture-of-Experts language model.
  • Training Data: Utilizes the lingshu-medical-mllm/ReasonMed dataset, which contains medical reasoning Q&A examples with multi-step rationales.
  • Safety Alignment: Incorporates safety tuning to prefer cautious, educational responses, avoid unsupported medical claims, and recommend professional consultation.

Intended Use Cases (Research & Development Only)

  • Biomedical research question answering
  • Medical exam-style reasoning and benchmark evaluation
  • Clinical knowledge retrieval experiments
  • Differential diagnosis reasoning research
  • Medical education support and safety-aligned medical AI research

Performance Highlights

  • MedQA: Achieves 67, outperforming the base gpt-20b at 62.
  • HealthBench: Scores 42.5, matching the base gpt-20b.

Important Limitations & Safety

This model is for research and development use only and is not intended for direct clinical diagnosis, treatment decisions, or patient management. Outputs may contain inaccuracies or hallucinations and require independent verification by qualified medical professionals. Users must adhere to strict ethical guidelines, biosafety, and biosecurity standards, and avoid any harmful applications.

Finetunes
1 models