yuvalkansal/QwQ-Med-3

TEXT GENERATIONConcurrency Cost:2Model Size:32.8BQuant:FP8Ctx Length:32kPublished:Mar 6, 2026Architecture:Transformer0.0K Cold

QwQ-Med-3 is a 32.8 billion parameter medical reasoning model developed by Yuval Kansal, fine-tuned from Qwen/QwQ-32B. It specializes in multi-hop reasoning paths derived from medical Knowledge Graphs, enabling it to provide structured, evidence-aligned chain-of-thought responses. This model is primarily designed for medical question answering and evaluation on ICD-coded clinical vignettes.

Loading preview...

QwQ-Med-3: Medical Reasoning Model

QwQ-Med-3 is a 32.8 billion parameter language model developed by Yuval Kansal, Bhishma Dedhia, and Niraj K. Jha, specifically engineered for advanced medical reasoning. It is fine-tuned from the Qwen/QwQ-32B base model using a Supervised Fine-Tuning (SFT) pipeline.

Key Capabilities

  • Multi-hop Medical Reasoning: The model is trained on question-answer pairs grounded in multi-hop reasoning paths extracted from a medical Knowledge Graph. This process teaches it to generate structured, evidence-based chain-of-thought reasoning.
  • Knowledge Graph Alignment: Its training methodology ensures that its reasoning is aligned with and supported by medical knowledge graph evidence, enhancing reliability in medical contexts.

Intended Use Cases

  • Medical Question Answering: Excels in scenarios requiring complex, multi-hop reasoning to answer medical queries.
  • Clinical Vignette Evaluation: Suitable for evaluating performance on ICD-coded clinical vignettes.
  • Research: Valuable for research into knowledge graph-guided language model training and domain-specific AI development.

This model is introduced in the paper "Bottom-up Domain-specific Superintelligence: A Reliable Knowledge Graph is What We Need" (arXiv:2507.13966). The associated code is available at jha-lab/bottom-up-superintelligence, and a dedicated benchmark, yuvalkansal/KG-Med-Bench, is provided for evaluation.