On June 9th, XUNFEIHEALTH (02506) officially launched the Spark Medical Large Model V3.5, which was trained on a fully domestic computing power platform. This upgrade focuses on two core, high-demand areas: clinical diagnosis and treatment, and public health management, achieving breakthrough progress in fields such as multimodal fusion, efficient long-text reasoning, and evidence-based medical decision-making. In key tasks including medical knowledge Q&A, medical language understanding, diagnostic and treatment recommendations, medical document generation, medical multimodal interaction, and medical multi-turn dialogue, the model's comprehensive performance has surpassed leading mainstream large models both domestically and internationally, such as GPT-5.5, DeepSeek-V4-Pro, and Qwen3.5-Max. It has also secured the top position in evaluations by authoritative institutions like IDC and the Shanghai Artificial Intelligence Laboratory.
This iteration not only realizes a multimodal technological innovation that deeply integrates speech, imaging, and semantics under the support of fully domestic computing power but also comprehensively enhances the capabilities of intelligent agents in medical scenarios. It enables multiple core applications for both doctors and the public to achieve a breakthrough from 0 to 1, taking the lead in crossing the chasm from technological exploration to large-scale practical application.
Comprehensive Enhancement of Core Capabilities and Leading Performance in Authoritative Evaluations
This version of the Spark Medical Large Model V3.5 has undergone a full-dimensional iteration from the technical foundation to the application layer, achieving breakthroughs in three core technologies and continuously strengthening the industry's technical barriers. First, the multimodal capability, which deeply integrates medical speech recognition, imaging-assisted diagnosis, and medical semantic understanding, has been comprehensively upgraded. It is the first in the industry to overcome the practical threshold for far-field multi-speaker voice recognition and automatic medical record generation in real clinical settings. Building upon a foundation where imaging quality control already surpasses the level of expert consultations, it deeply combines imaging recognition with diagnostic reasoning, with the imaging report generation capability crossing the practical threshold for the first time in the industry.
Second, it is the first to achieve efficient long-text training using DSA (Dynamic Sparse Attention) and MTP (Multi-Token Prediction) on a domestic computing power platform, increasing inference throughput in medical long-context scenarios by 4.5 times. This provides autonomous and controllable underlying technical support for the localized model deployment and efficient iteration in hospitals and regional healthcare systems.
Third, relying on the algorithmic breakthroughs of the Spark Medical Large Model V3.5, high-quality medical datasets, and reinforcement learning with evidence-based thinking chains, the capabilities of intelligent agents for medical scenarios have been comprehensively enhanced. The overall performance and practical utility of the various product series built on this foundation have achieved a leapfrog improvement.
The impressive technical performance is attributed to the synergistic optimization achieved by the integration of high-quality medical knowledge data and algorithmic innovation. In terms of high-quality medical data, cross-modal alignment of data such as speech, imaging, and medical record text has been achieved. The model is continuously trained using 1.6 billion de-identified medical speech records and 1.2 billion real-world diagnostic and treatment records. The platform adds over 2.2 million real medical data entries daily, including speech, imaging, and medical records, for learning. Through the continuous accumulation of high-quality data, it has built an industry-unique data flywheel and a barrier to autonomous evolution.
In terms of algorithmic innovation, based on high-quality medical data and weakly supervised data from doctor interaction feedback, an evidence-based reasoning technology for complex tasks—'evidence alignment, reflection verification, expert reinforcement'—has been realized, significantly improving the accuracy of the medical large model.
Multiple authoritative evaluations confirm the model's technological leadership. According to the authoritative assessment in IDC's "China Healthcare Large Model Technology Assessment, 2026," the comprehensive strength of the Xunfei Spark Medical Large Model ranks first in the industry. This evaluation covered 15 core indicators, with the model securing first place in 12 capabilities including assisted diagnosis, medical record generation, health consultation, post-diagnosis management, and follow-up. The test results indicate that the model performs best in product viability, safety, and large-scale application capability, fully confirming that its technical strength and scenario adaptability are at the industry's leading level.
In the latest intelligent agent evaluation by the Shanghai Artificial Intelligence Laboratory's MedBench, the model topped the list with a comprehensive score of 98.9, ranking first in multiple core capabilities such as medical scenario perception and interaction, medical multi-agent collaboration, and medical safety, ethics, and compliance.
Deepening Scenario Application, Advancing in Both Clinical Diagnosis and Health Management
Leveraging its powerful underlying model capabilities, the Spark Medical Large Model V3.5 comprehensively empowers the AI Diagnosis Assistant and AI Health Assistant, enabling the deep release of technological value. In the medical record writing scenario, the multimodal medical record generation technology has been upgraded across all dimensions, breaking through multiple core technologies including multi-source doctor-patient far-field recognition, multi-speaker directional voice enhancement, and a non-autoregressive architecture for medical speech recognition. The end-to-end medical record generation rationality rate has reached 80%. In real-world applications at several top-tier tertiary hospitals, the doctor adoption rate for generated records reached 91%, and record writing time was reduced by 52%. Medical speech recognition and medical record generation have crossed the practical threshold for the first time in the industry.
In medical imaging report generation and quality control scenarios, through upgrades in core technologies such as multi-sequence joint spatial modeling and the integration of diagnostic reasoning capabilities, the ability to generate complex imaging reports and perform quality control has been further enhanced. The doctor adoption rate for generated X-ray and MR reports has reached 75%, leading the industry in crossing the practical threshold.
For clinical doctors, empowered by evidence-based long-chain reasoning technology, the diagnosis and treatment assistant intelligent agent, centered on authoritative cutting-edge medical knowledge services in Chinese and English and evidence-based reasoning, has been comprehensively upgraded. It can provide clinical decision support with multi-step reasoning, deep reflection, and complete evidence traceability. In other essential diagnostic and treatment scenarios for doctors, such as medical record content quality control, intelligent medication review, and general and cross-specialty assisted diagnosis, the model's various capabilities have been further enhanced, maintaining its industry leadership.
For the general public, the model has simultaneously upgraded technologies such as multimodal consultation, dynamic memory management of health records, and precise personalized consultation, comprehensively empowering the personal health assistant. It continues to maintain industry-leading performance in tasks such as interpreting lab and test reports, interpreting physical examination reports, medication inquiry (by photographing medicine boxes), and multi-turn proactive consultation.
Technological Upgrade Empowers the Entire Business Chain, Fortifying the Foundation for Long-Term Growth
The development of medical AI is not a short-term race but a long-term contest concerning safety, compliance, professionalism, and ecosystem. The stringent requirements in the medical field for AI to have low hallucination rates, be explainable, traceable, and highly compliant determine that generic internet data cannot support the continuous evolution of professional medical models. Only by being rooted in frontline diagnosis and treatment, serving hundreds of millions of residents, and accumulating native scenarios and practical experience can medical AI have a solid foundation for steady and long-term progress.
As a pioneer in the innovative practice of medical artificial intelligence, XUNFEIHEALTH has always adhered to a long-term development philosophy. It has taken the lead in completing a full-scenario strategic layout from primary care to tertiary hospitals, and from diagnosis and treatment to health, establishing a comprehensive service ecosystem that covers the nation with multi-dimensional complementarity. At the primary care end, it covers 806 districts and counties nationwide and over 77,000 primary care institutions, cumulatively assisting in over 1.2 billion diagnoses. In regional healthcare, benchmark projects like Imaging Cloud and the digital family doctor platform have been implemented in multiple regions, breaking down regional medical data barriers. At the hospital end, it collaborates with over 600 graded hospitals, including more than 50 of China's top 100 hospitals and 7 of the top 10 hospitals. On the resident end, the Xunfei Xiaoyi APP has surpassed 30 million cumulative downloads, completing over 180 million health consultations.
Comprehensive regional layout, full-scenario services, and massive practical experience collectively build XUNFEIHEALTH's multi-layered, difficult-to-replicate comprehensive barriers.
The upgrade and release of the Spark Medical Large Model V3.5 is a concentrated reflection and a milestone achievement of XUNFEIHEALTH's years of practice in the virtuous cycle of "deep technological cultivation + scenario implementation." Relying on a long-honed technological foundation, full-scenario implementation capability, and a mature service system, the model has achieved an all-around leap from its underlying architecture to application capabilities. The upgraded technological capabilities will, in turn, permeate all links of the chain, including primary care diagnosis and treatment, tertiary hospital services, and resident health management, promoting deep synergy and a spiral ascent of technology, scenarios, and the ecosystem.
Riding the industry tailwind of artificial intelligence empowering healthcare, the value of the Spark Medical Large Model V3.5 will continue to be unleashed, driving the data flywheel to operate at high speed and continuously expanding the industry's leading advantage. Looking to the future, XUNFEIHEALTH will adhere to the development philosophy of "reaching for the sky with technology and grounding with application." With the Spark Medical Large Model as the core foundation, it will continue to tackle technological innovation, deepen large-scale scenario implementation, help improve the quality and efficiency of medical services, and safeguard the health of all people.