Henrychur/DiagAgent-14B

TEXT GENERATIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 16, 2025License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

DiagAgent-14B by Henrychur is a 14.8 billion parameter large language model optimized for interactive, multi-turn diagnostic reasoning in medical contexts, featuring a 32768 token context length. It is specifically trained using reinforcement learning (GRPO) within a virtual clinical environment (DiagGym) to recommend informative examinations, update diagnoses, and commit to final diagnoses. This model excels in medical diagnostic accuracy and examination recommendation, outperforming general-purpose LLMs and other agentic systems in end-to-end clinical reasoning tasks.

Loading preview...

DiagAgent-14B: RL-Optimized Diagnostic Agent

DiagAgent-14B is a 14.8 billion parameter large language model developed by Henrychur, specifically optimized for interactive, multi-turn diagnostic reasoning in medical applications. Unlike traditional one-shot medical LLMs, this model is designed to simulate a clinical reasoning process, making it highly effective for complex diagnostic scenarios.

Key Capabilities

  • Interactive Diagnostic Reasoning: Recommends informative examinations, updates working diagnoses as new evidence emerges, and decides when to finalize a diagnosis.
  • Reinforcement Learning Optimization: Trained end-to-end using multi-turn Reinforcement Learning (GRPO) within the DiagGym virtual clinical environment, enabling safe and closed-loop learning.
  • Superior Performance in Medical Diagnostics: Achieves a Diagnosis Accuracy of 86.73% in single-turn evaluations and an F1 score of 47.89% with 61.63% Accuracy in end-to-end evaluations, significantly outperforming larger general-purpose LLMs like GPT-4o, Claude-4-sonnet, and Qwen2.5, as well as other agentic systems.
  • Efficient Resource Utilization: Despite its specialized capabilities, it can be run locally using transformers or served with vLLM, with tips provided for VRAM constrained environments.

Good For

  • Medical AI Assistants: Developing AI systems that can engage in multi-turn dialogues for patient diagnosis.
  • Clinical Decision Support: Assisting healthcare professionals by suggesting relevant tests and refining diagnoses based on new information.
  • Medical Education and Training: Simulating diagnostic scenarios for training purposes in a risk-free virtual environment.
  • Research in AI for Medicine: Providing a robust baseline and framework for further research into agentic medical AI systems. Details on the model's training and evaluation can be found in the associated paper: https://arxiv.org/abs/2510.24654.

Finetunes
1 models