TachyHealth/Gazal-R1-32B-GRPO-preview

TEXT GENERATIONConcurrency Cost:2Model Size:32BQuant:FP8Ctx Length:32kPublished:May 26, 2025License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Gazal-R1-32B-GRPO-preview is a 32.8 billion parameter causal language model developed by TachyHealth, built upon Qwen 3 32B. It is specifically designed and fine-tuned for medical reasoning and clinical decision-making, leveraging a two-stage training pipeline including Group Relative Policy Optimization (GRPO). This model excels at diagnostic reasoning, treatment planning, and prognostic assessment, achieving state-of-the-art performance on medical benchmarks like MedQA and MMLU Pro (Medical).

Loading preview...

Gazal-R1-32B-GRPO-preview: Medical Reasoning Specialist

Gazal-R1 is a 32.8 billion parameter language model from TachyHealth, built on Qwen 3 32B, and specifically engineered for advanced medical reasoning. It utilizes a two-stage training approach, combining Supervised Fine-Tuning (SFT) with Group Relative Policy Optimization (GRPO) for alignment. This methodology enables the model to provide structured clinical thinking, including step-by-step explanations within <think> tags.

Key Capabilities & Performance

  • Medical Expertise: Specialized training on over 107,000 synthetic medical reasoning examples, covering diagnostic reasoning, treatment planning, and prognostic assessment.
  • State-of-the-Art Benchmarks: Achieves 87.1% on MedQA, 81.6% on MMLU Pro (Medical), and 79.6% on PubMedQA, outperforming models up to 12 times larger (e.g., Llama 3.1 405B Instruct) in medical domains.
  • Transparent Reasoning: Generates structured explanations following established clinical reasoning frameworks.
  • Parameter Efficiency: Employs advanced techniques like Weight-Decomposed LoRA (DoRA) and Rank-Stabilized LoRA (rsLoRA) for efficient fine-tuning.
  • Context Length: Supports a native context length of 32,768 tokens, extensible to 131,072 with YaRN.

Ideal Use Cases

  • Research and Education: Excellent for medical education, clinical reasoning research, and academic medical writing assistance.
  • Professional Support (Supervised): Can aid in literature review, clinical case analysis, and medical documentation, always requiring verification by qualified professionals.

Important Disclaimers

Gazal-R1 is a research model and NOT intended for direct clinical use, diagnosis, or treatment. All outputs must be independently verified by medical professionals. It has a knowledge cutoff and may hallucinate; it should never be used for emergency medical situations.