Healthcare and legal large models break through: From lab to clinic and courtroom
In September 2026, multiple domestic AI vendors, in collaboration with medical institutions, released vertical large models for clinical decision support. These models underwent real-world testing in scenarios such as rare disease screening, medical image report interpretation, and medication risk alerts, and received filing approval from health departments in several provinces and cities. Unlike general-purpose large models, medical large models incorporate large volumes of de-identified electronic medical records, public clinical trial data, and treatment guidelines during training, and undergo multiple rounds of adversarial validation to ensure outputs align with medical logic. A top-tier tertiary hospital publicly reported that the model's sensitivity in identifying early-stage lung cancer approached that of senior radiologists, but emphasized that manual review remains necessary. Meanwhile, significant progress was made in the legal domain. A legal tech company, in partnership with several law firms, launched a judicial large model that completed the mock exam for the lawyer qualification test in September 2026, achieving over 85% accuracy on basic legal knowledge questions. It also demonstrated practical value in tasks such as contract clause risk identification and similar case retrieval. The model has been piloted in pre-litigation mediation assistance systems in some grassroots courts, primarily handling automatic document generation and similar case recommendations to reduce the workload of court clerks.
Low-code Agent building platforms launch en masse: Enabling non-technical users to create AI assistants
In September 2026, more than ten cloud computing and AI platforms jointly launched low-code Agent building tools. The core of these tools is to enable business users to quickly generate intelligent Agents that can call external APIs, databases, and knowledge bases by dragging and dropping process nodes and configuring natural language instructions. For example, a major e-commerce platform launched a 'product operation Agent' low-code tool that allows operations staff to describe promotion rules in a single sentence, and the system automatically generates corresponding code and logic flows, directly connecting to inventory systems and customer service channels. Another open-source framework provider introduced a 'zero-code Agent workshop' that supports users in orchestrating steps on a canvas and debugging in real time. Industry observers noted that the maturation of low-code Agents will accelerate the vision of 'everyone can build their own AI assistant' within enterprises, but it also brings new challenges in security control and permission governance.
On-device model updates: Smart devices and in-vehicle scenarios see offline intelligence upgrades
In September 2026, multiple chip and device manufacturers jointly released a new generation of on-device large model updates. The update focused on three main scenarios: smart homes, wearable devices, and smart vehicles. In the smart home domain, the new model improves local voice wake-up word recognition accuracy to over 99%, and supports executing complex commands offline (such as 'find last week's saved recipe and read the steps') with latency under 200 milliseconds. A leading home appliance brand announced that its 2027 smart speakers will come standard with this model and allow users to customize wake-up words via their phones. In in-vehicle scenarios, the on-device model enables multimodal interaction during driving: passengers can describe destination characteristics in natural language (e.g., 'find a quiet and affordable Sichuan restaurant'), and the model combines local map data with user history preferences to complete inference entirely on the vehicle's system without requiring an internet connection. Additionally, some car models have integrated driving behavior analysis models that can identify fatigue driving offline and issue timely voice alerts.
Industry observation: A key shift from model capability to engineering deployment
Looking at the multiple developments in September 2026, the AI industry is shifting from 'benchmark scores and parameter counts' to 'deployment and real-world results.' The regulatory approval and trial licensing of medical and legal large models mark an important step for AI in serious domains. The concentrated release of low-code Agent tools reflects the industry's efforts to lower technical barriers and expand application scenarios. The functional updates to on-device models further validate the advantages of offline intelligence in privacy protection and real-time performance. It is foreseeable that as vertical industry standards are gradually established and the Agent ecosystem matures, the practical value of large models will be more fully realized in the next six months. However, security compliance, model bias, and long-term maintenance costs remain real challenges for developers.