At the 2026 World Artificial Intelligence Conference in Shanghai from July 17th to 20th, JD Health (JD HEALTH) showcased its latest medical AI advancements. The company unveiled its core technical proposition of "Determinism from Execution" and announced a major upgrade to its JoyAI-MedCode (Jingyi Qianxun) codified reasoning system. Furthermore, JD Health announced a collaboration with the Shanghai Artificial Intelligence Laboratory to jointly establish the MedBench v5.0 industry evaluation benchmark. The company's two flagship AI products, the AI Doctor "Dawei" and "JD Zhiyi," were also exhibited at the event.
Core Technology Advancement
During the forum, JD Health formally announced the significant v2.0 upgrade to the core reasoning engine of Jingyi Qianxun, JoyAI-MedCode. This upgrade is positioned to propel medical AI from traditional "natural language reasoning" towards "verifiable code-based reasoning."
Executives from JD Health's medical AI division stated that the next phase of competition will not center on which model has the most parameters, but on which can deliver AI to doctors and users reliably, trustworthily, and at scale. Unlike general-purpose large language models that rely on probabilistic generation with opaque reasoning processes and difficult-to-verify conclusions, JoyAI-MedCode v2.0 is built on the principle of "guidelines as code." This approach compiles authoritative clinical guidelines, such as those from NCCN and CSCO, into executable decision-making code. Diagnostic and treatment suggestions are generated directly by running this code, with guideline pathways corresponding to logical code branches. Every conclusion can be precisely traced back to the relevant section of the source guideline and can be batch-validated akin to software unit testing, fundamentally ensuring the determinism, compliance, and traceability of medical reasoning.
Establishing an Industry Standard
At WAIC, JD Health and the Shanghai Artificial Intelligence Laboratory officially announced their partnership to co-develop MedBench v5.0. This initiative integrates their self-developed medical AI evaluation system into a nationally recognized authoritative benchmark, aiming to create a unified standard for assessing medical AI capabilities across the industry.
The evaluation capabilities incorporated into MedBench v5.0 cover three key areas: multimodal online consultations, the inclusion of a 3D imaging evaluation dataset into the main benchmark for full-modal large models, and the integration of an evidence-based evaluation dataset into a specialized track. Two of its core underlying datasets are noted for addressing industry gaps. The MedRealMM multimodal real-scenario evaluation set contains thousands of ethically reviewed real-world clinical multimodal cases, providing the first systematic empirical evidence of the significant impact of image information on consultation evaluations, with performance gaps between image-based and text-only evaluations reaching 20-30 points. The Med3DVQA 3D imaging evaluation set is constructed from over 20,000 complete CT/MRI examinations and approximately 500,000 radiology report Q&A pairs. It covers seven types of visual question-answering tasks and is positioned as China's first fully open-source 3D evaluation set for complete radiological examinations, addressing the industry's lack of evaluation benchmarks for cross-temporal and cross-sequence reasoning in 3D imaging.
Showcasing AI Products
At the WAIC exhibition, JD Health showcased its two core AI products. The AI Doctor "Dawei," designed for consumer users, is built on the Jingyi Qianxun medical large language model. It deeply integrates capabilities across the entire healthcare service chain, including online consultations, at-home rapid testing, nurse home visits, and medication fulfillment, covering over a thousand conditions in both Western and Traditional Chinese Medicine. "JD Zhiyi," developed for medical professionals, is an evidence-based AI tool tailored for clinical workers by JD Health. It deeply integrates access to over 40 million medical literature articles, 40,000 clinical guidelines, and more than 30,000 drug package inserts.
Comments