jdh-algo/JoyMed-32B-v1.0

JoyMed-32B-v1.0

JoyMed: A Leading Medical Foundation Model with Adaptive Reasoning

 

Ju Huang  Xinyi Liu  Sheng Shi  Fangru Zhou  Jun Zhao  Jun Xu 

JDH Algo, JD Health International Inc. 

🔥 News

• [2026-04-02]: 🎉 We are pleased to announce the release of 🤗 JoyMed-32B-v1.0 ! Welcome to use and explore!
• [2026-03-27]: 📝 We open-source the evaluation script for the related benchmark today, please refer to evaluation/Readme.md!
• [2026-03-26]: 📚 We release 📂 MedDocBench, a self-collected medical benchmark dataset to facilitate research and evaluation in medical multimodal learning!
• [2026-03-26]: ⚙️ We have released 🤗 JoyMed-8B-v1.0 for research and application! Welcome to use and explore!
• [2026-03-26]: 🎉 We propose JoyMed, a medical foundation model with adaptive reasoning that balances reasoning accuracy and efficiency, achieves SOTA across multiple benchmarks, and advances the translation of medical multimodal large models to clinical applications!

📖 Overview

Reasoning capabilities are foundational to medical multimodal large models (MMLMs), as they enable trustworthy diagnosis, interpretable decision-making, and effective management of complex clinical cases. However, mainstream MMLMs either lack explicit reasoning capacities or rely on fixed, end-to-end, undifferentiated mandatory reasoning paradigms. This not only introduces chain-of-thought redundancy and wastes computational resources but also degrades performance in pure perceptual tasks due to unnecessary reasoning overhead. The core bottleneck lies in the inherent trade-off between reasoning accuracy and computational efficiency: thorough reasoning ensures rigor for complex tasks but imposes redundant costs on simple tasks, whereas efficiency-oriented simplified output paradigms lack the sufficient reasoning completeness critical for complex clinical decision-making. To address these challenges, we propose JoyMed, a leading medical foundation model with adaptive reasoning. Building upon direct output and reasoning-augmented output paradigms, we introduce an adaptive reasoning mechanism that directly outputs results for trivially simple tasks to optimize efficiency, while generating stepwise reasoning traces for complex tasks to balance accuracy and interpretability. Experimental results demonstrate that JoyMed achieves state-of-the-art performance across multiple benchmarks, which effectively balances the core clinical requirements of comprehensive accuracy and efficient result acquisition, marking an exploratory step toward translating MMLMs from laboratory research to practical clinical applications.

Key Features

• Superior Performance: Our model achieves leading performance across multiple core medical benchmarks,encompassing medical image understanding, text-based question answering, medical document comprehension, and medical report generation, demonstrating its robust capabilities. This outstanding performance stems from our carefully designed two-stage training strategy. First, fine-grained vision-language alignment training significantly enhances the model’s perceptual ability for key regions such as lesions and anatomical structures. Subsequently, reinforcement on complex tasks like report generation and case analysis establishes precise associations between visual regions and textual descriptions, providing solid support for high-level medical visual understanding and question answering.
• Adaptive Reasoning: To strike an optimal balance between computational efficiency and deep reasoning, we innovatively propose an adaptive reasoning mechanism. Its core involves constructing a difficulty-tiered dataset and designing corresponding training strategies to mitigate potential mode collapse during the model’s autonomous reasoning process, enabling independent assessment of the intrinsic complexity of problems. Our proposed model operates in three modes: direct output, chain-of-thought reasoning, and adaptive thinking. This design allows the model to intelligently allocate computational resources based on task demands while maintaining high accuracy and interpretability, thereby achieving an optimal trade-off between effectiveness and efficiency.

🚀 Model Zoo

JoyMed comes in two variants with different parameter configurations:

Model	Parameters	Hugging Face
JoyMed-8B-v1.0	8B	🤗 JoyMed-8B-v1.0
JoyMed-32B-v1.0	32B	🤗 JoyMed-32B-v1.0

🏆 Performance

Medical Textual Question Answering Benchmarks

The best results on each benchmark and average accuracy are highlighted in bold, and the scores with underline indicate the second best. Note that MedXQA and SGPQA denote MedXpertQA-Text and SuperGPQA-medical benchmarks.

Model	PubMedQA	MedMCQA	MedXQA	CMMLU	SGPQA	MedQA (USMLE)	MedQA (MCMLE)	Medbullets (op-4)	Medbullets (op-5)	Avg.
Proprietary Models
GPT 4.1	76.00	87.98	30.82	81.02	50.60	77.07	81.73	78.90	73.38	70.83
GPT 5	78.00	62.99	40.75	82.93	49.54	76.96	74.00	88.93	87.30	71.27
Doubao Seed 1.6	76.00	75.06	30.67	91.67	55.19	93.48	94.02	82.79	76.62	75.06
Open-Source Models (<10B)
MedGemma 4B	73.00	52.26	13.10	43.96	21.52	55.54	41.10	48.05	42.53	43.45
Qwen3-VL 8B	73.20	60.05	14.98	79.07	35.68	65.67	85.61	55.84	48.70	57.64
HuatuoGPT-V 7B	73.60	51.95	10.33	71.12	22.11	52.95	73.09	43.51	37.66	48.48
Lingshu 7B	75.40	56.13	16.45	69.02	27.51	63.39	75.98	62.66	52.92	55.50
Citrus-V 8B	74.80	55.10	16.90	71.19	29.47	64.89	76.94	59.09	54.22	55.84
Hulu-Med 7B	77.20	67.51	18.53	71.72	31.10	73.45	78.93	64.94	57.47	60.09
JoyMed 8B	78.20	65.36	23.67	82.75	37.10	82.64	92.06	73.05	68.18	67.00
JoyMed 8B auto	79.40	66.58	23.55	83.05	38.04	84.84	91.42	74.35	70.46	67.96
JoyMed 8B thinking	79.00	66.89	24.37	83.35	39.78	85.23	91.77	73.38	70.13	68.21
Open-Source Models (>10B)
MedGemma 27B	79.00	63.23	22.01	60.24	33.18	81.15	64.89	67.86	65.58	59.68
Qwen3-VL 32B	72.00	69.57	18.65	86.87	46.24	77.77	89.73	63.96	53.90	64.30
HealthGPT 14B	69.40	63.33	12.45	55.36	25.59	66.93	52.83	53.57	50.00	49.94
HealthGPT 32B	74.20	64.04	13.84	69.47	35.43	68.89	68.86	50.65	46.43	54.65
HuatuoGPT-V 34B	71.00	55.08	12.20	77.64	28.06	58.52	76.09	44.81	39.29	51.41
Lingshu 32B	78.20	65.05	22.86	82.37	40.80	74.94	86.98	68.51	63.31	64.78
Citrus-V 33B	78.40	65.62	22.20	83.27	41.63	80.28	87.65	67.53	66.23	65.87
Hulu-Med 14B	79.80	70.28	23.67	75.62	37.75	78.48	80.88	70.45	67.86	64.98
Hulu-Med 32B	80.20	72.56	23.92	76.07	41.71	80.13	84.30	71.75	68.18	66.54
JoyMed 32B	80.20	72.72	32.65	88.22	48.93	87.43	93.90	80.52	76.30	73.43
JoyMed 32B auto	80.00	72.96	31.35	88.30	48.93	90.34	94.31	81.82	78.57	74.06
JoyMed 32B thinking	78.60	73.39	31.31	89.27	51.25	89.16	94.31	80.84	76.62	73.86

Medical Visual Question Answering Benchmarks

The best results on each benchmark and average accuracy are highlighted in bold, and the scores with underline indicate the second best. Note that MedXQA and GMAI-MMB denote MedXpertQA-mm and GMAI-MMBench-test benchmarks.

Model	VQA-RAD	MedXQA	SLAKE	PATH-VQA	PMC-VQA	OmniMedVQA	GMAI-MMB	Avg.
Proprietary Models
GPT 4.1	62.53	43.35	72.54	54.97	38.76	55.14	58.52	55.12
GPT 5	68.37	51.48	65.82	31.74	36.10	38.44	56.18	49.73
Doubao Seed 1.6	33.49	45.75	67.28	47.58	49.94	61.68	48.50	50.60
Open-Source Models (<10B)
MedGemma 4B	72.06	22.05	78.32	48.64	48.02	70.04	45.59	54.96
Qwen3-VL 7B	63.41	25.00	72.11	43.65	54.01	76.90	54.31	55.63
HuatuoGPT-V 7B	67.85	22.30	69.39	44.29	53.84	75.14	51.56	54.91
Lingshu 7B	68.74	26.90	82.90	60.23	55.77	82.41	54.02	61.57
Citrus-V 8B	64.30	25.10	84.91	62.00	55.64	72.69	45.43	57.45
Hulu-Med 7B	74.50	27.70	82.66	62.57	66.95	83.70	54.28	64.62
JoyMed 8B	75.83	32.60	86.53	74.16	57.19	82.36	59.85	66.93
JoyMed 8B auto	76.50	33.25	87.97	75.06	58.34	81.47	60.37	67.56
JoyMed 8B thinking	79.16	33.20	86.82	75.34	58.52	81.43	60.35	67.83
Open-Source Models (>10B)
MedGemma 27B	63.86	33.10	76.17	47.60	45.35	59.78	40.59	52.35
Qwen3-VL 32B	68.96	30.15	77.41	49.78	57.28	77.90	55.86	59.62
Lingshu 32B	75.39	31.00	87.68	64.76	57.23	82.95	55.32	63.21
HealthGPT 14B	64.08	24.55	67.43	58.67	56.90	76.45	45.67	56.25
HealthGPT 32B	64.75	26.40	70.58	62.93	54.93	73.09	46.61	57.04
HuatuoGPT-V 34B	63.64	22.65	73.02	44.92	56.79	73.93	54.29	55.61
Citrus-V 33B	77.83	29.15	88.40	63.89	59.74	77.02	53.50	64.22
Hulu-Med 14B	74.70	29.65	84.24	64.55	68.80	85.22	57.42	66.37
Hulu-Med 32B	79.60	33.40	85.86	66.44	69.52	84.41	59.60	68.40
JoyMed 33B	88.69	41.25	92.93	91.29	61.46	84.75	61.02	74.48
JoyMed 33B auto	89.14	42.45	94.70	92.19	61.17	84.22	61.21	75.01
JoyMed 33B thinking	87.81	41.00	93.27	89.49	61.32	81.31	60.62	73.54

Medical Document Understanding Benchmarks

The best results on each benchmark and average accuracy are highlighted in bold, and the scores with underline indicate the second best.

	Laboratory Test Report			GMD
Model	abnormalityQA	fullparsing	simpleQA	Simple QA	Complex QA	Avg.
Proprietary Models
GPT 4.1	45.39	73.97	66.00	45.60	55.85	57.36
GPT 5	71.87	71.41	78.75	59.35	57.04	67.68
Doubao Seed 1.6	82.16	81.09	85.00	73.60	79.25	80.22
Open-Source Models (<10B)
MedGemma 4B	13.31	36.39	17.75	18.10	20.05	21.12
Qwen3-VL 8B	48.17	79.71	84.50	78.72	79.05	74.03
HuatuoGPT-V 7B	7.54	32.14	9.50	17.40	9.60	15.24
Lingshu 7B	29.50	62.70	70.25	60.47	53.70	55.32
Citrus-V 8B	91.16	92.57	92.57	83.22	89.45	90.18
Hulu-Med 7B	11.35	43.07	19.75	19.30	15.72	21.84
JoyMed 8B	88.99	93.39	92.00	83.47	86.72	88.91
JoyMed 8B auto	90.80	88.87	94.00	85.20	85.67	88.91
JoyMed 8B thinking	90.80	88.87	94.00	85.10	85.47	88.85
Open-Source Models (>10B)
MedGemma 27B	10.89	35.15	28.25	19.50	14.55	21.67
Qwen3-VL 32B	50.86	83.85	94.50	81.00	84.90	79.02
Lingshu 32B	35.34	71.19	74.00	62.00	58.38	60.18
HealthGPT 14B	5.40	39.37	13.00	20.00	9.20	19.26
HealthGPT 32B	6.12	33.91	21.75	21.10	9.97	20.59
HuatuoGPT-V 34B	7.81	32.78	10.75	19.00	8.00	17.11
Citrus-V 33B	92.45	92.62	97.75	84.30	83.78	90.18
Hulu-Med 14B	7.59	45.02	26.75	22.40	15.40	26.45
Hulu-Med 32B	11.45	41.32	31.50	30.30	18.90	28.09
JoyMed 32B	90.12	93.68	91.50	84.30	85.96	91.77
JoyMed 32B auto	94.65	94.67	94.00	87.50	85.85	94.44
JoyMed 32B thinking	94.58	94.46	94.50	87.40	81.40	94.51

Medical Image Report Generation Benchmarks

The best results on each benchmark and average accuracy are highlighted in bold, and the scores with underline indicate the second best.

	CheXpert Plus		IU XRAY
Model	ROUGE-L	RaTE	ROUGE-L	RaTE
Proprietary Models
GPT 4.1	24.50	45.50	32.63	50.91
GPT 5	24.48	51.26	31.72	56.64
Doubao Seed 1.6	19.27	45.49	22.67	53.76
Open-Source Models (<10B)
MedGemma 4B	26.01	51.23	39.51	61.99
Qwen3-VL 7B	21.64	46.51	25.22	52.88
HuatuoGPT-V 7B	21.40	46.58	29.96	54.91
Lingshu 7B	26.50	45.40	44.52	60.30
Citrus-V 8B	28.94	51.07	24.78	56.25
Hulu-Med 7B	28.94	51.07	36.15	63.50
JoyMed 8B	32.54	55.68	42.85	64.90
JoyMed 8B auto	31.68	55.73	44.94	65.96
JoyMed 8B thinking	31.99	55.89	44.40	65.70
Open-Source Models (>10B)
MedGemma 27B	17.65	48.73	32.59	58.70
Qwen3-VL 32B	19.28	47.60	25.96	59.12
Lingshu 32B	25.29	46.18	45.06	65.22
HealthGPT 14B	21.29	47.82	23.89	52.33
HealthGPT 32B	12.50	45.15	12.74	45.75
HuatuoGPT-V 34B	23.97	45.51	29.12	57.87
Citrus-V 33B	29.58	52.45	45.66	64.74
Hulu-Med 14B	23.08	50.25	36.15	63.50
Hulu-Med 32B	29.20	52.87	47.02	65.43
JoyMed 32B	30.79	54.33	42.19	63.85
JoyMed 32B auto	28.59	54.28	40.63	63.14
JoyMed 32B thinking	30.04	53.60	44.05	66.19

🛠️ Installation

1. Installing vLLM

   uv venv
   source .venv/bin/activate
   uv pip install -U vllm --torch-backend=auto
   
   # Update transformers to support the latest models.
   pip install -U "transformers==5.2.*"
   

2. Running JoyMed-8B-v1.0

   vllm serve jdh-algo/JoyMed-8B-v1.0 \
       --tensor-parallel-size 8 \
       --mm-encoder-tp-mode data \
       --mm-processor-cache-type shm \
       --enable-prefix-caching \
       --trust-remote-code \
       --gpu-memory-utilization 0.9
   

💻 Quick Start

Instruct mode

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="EMPTY"
)

response = client.chat.completions.create(
    model="jdh-algo/JoyMed-8B-v1.0",
    messages=[{"role": "user", "content": "What are the common causes of hypertension in adults? /no_think"}] # end with '/no_think' or nothing
)

print(response.choices[0].message.content)

Thinking mode

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="EMPTY"
)

response = client.chat.completions.create(
    model="jdh-algo/JoyMed-8B-v1.0",
    messages=[{"role": "user", "content": "What are the common causes of hypertension in adults? /think"}]
)

print(response.choices[0].message.content)

Auto mode

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:8000/v1",
    api_key="EMPTY"
)

response = client.chat.completions.create(
    model="jdh-algo/JoyMed-8B-v1.0",
    messages=[{"role": "user", "content": "What are the common causes of hypertension in adults? /auto_think"}]
)

print(response.choices[0].message.content)

🏛 License

This project is licensed under the Apache License (Version 2.0). For models and datasets, please refer to the original resource page and follow the corresponding License.