On-device models see concentrated launches: Lightweight design and privacy protection become key selling points
Around September 2026, multiple AI vendors successively launched on-device large models targeting smartphones, PCs, and IoT devices. Compared to cloud-based models, these models typically have 1 to 3 billion parameters and achieve local operation through quantization, pruning, and distillation techniques. Industry insiders point out that the core advantage of such models is that data does not need to be uploaded to the cloud, and they can perform text summarization, intent recognition, and basic conversations offline, thereby meeting the stringent data privacy requirements of industries such as finance and healthcare.
For example, a leading internet company released a new-generation on-device language model in mid-September, claiming it can achieve an inference speed of 40 tokens per second on Snapdragon 8-series chips, with memory usage controlled within 1.5GB. Another chipmaker simultaneously updated its accompanying inference engine, supporting dynamic batching and operator fusion to further lower the barrier for on-device deployment. Meanwhile, multiple smartphone vendors began pre-installing on-device AI assistants in new models, offering localized voice notes, real-time translation, and photo album management features.
Notably, the combination of privacy-preserving computing and on-device models has become a new trend. Some vendors have introduced federated learning frameworks during model training, allowing users to fine-tune models locally and upload only gradient information instead of raw data, thereby enhancing personalization while remaining compliant. Analysts believe this direction could promote AI adoption in highly regulated scenarios such as government and finance.
Multimodal agent frameworks accelerate deployment: From dialogue to task execution
In September, multiple companies released or upgraded agent frameworks based on multimodal large models, extending AI capabilities from simple text-based dialogue to cross-application operations and complex task orchestration. These agents typically integrate vision, speech, and text understanding modules, and can perform processes such as ordering food, expense reporting, and data entry through screenshot recognition, API calls, or GUI operations.
For example, a domestic AI unicorn open-sourced a multimodal agent development kit in mid-September, allowing developers to define workflows through natural language. The agent can autonomously invoke web browsers, code editors, and office software. In a demonstration, the agent, following user voice commands, automatically extracted tabular data from email attachments, called data analysis scripts to generate visual reports, and finally sent them to specified contacts via email. The entire process required no manual coding or interface configuration.
Another international tech company added an agent orchestration service to its cloud platform, allowing enterprises to connect internal systems (such as CRM, ERP) and third-party applications to agents via low-code methods. The service includes built-in security audit and permission control modules to ensure agents comply with corporate compliance requirements when accessing sensitive data. Industry analysis indicates that the maturity of multimodal agents is still limited by long-context processing capabilities and tool-calling stability, but the product updates in September show that engineering deployment has entered a practical stage.
Low-code AI development tools emerge: Lowering the barrier for small-to-medium enterprises
Along with the continuous improvement of model capabilities, tools that lower the barrier for developing AI applications also emerged in clusters in September. Multiple vendors launched AI workshop products targeting non-technical users, supporting the generation of customized AI applications through drag-and-drop interfaces, preset templates, and natural language descriptions. These tools typically come with built-in basic models for text classification, image recognition, and speech synthesis, allowing users to configure business logic without understanding model details.
For example, a cloud service provider launched an AI app building platform in early September, where users only need to upload a small amount of sample data, and the platform automatically completes model selection, hyperparameter tuning, and inference interface deployment. The platform also offers version management and A/B testing features to facilitate rapid iteration. Another startup focused on customer service scenarios, launching a zero-code intelligent chatbot builder that supports automatically extracting question-answer pairs from knowledge base documents and delivering accurate replies based on retrieval-augmented generation (RAG) technology.
For small-to-medium enterprises, such tools significantly reduce the human and financial costs of AI deployment. Tasks that previously required assembling an algorithm team can now be directly handled by business personnel. However, the industry also warns that low-code tools may perform poorly in long-tail, complex scenarios, and enterprises still need to select based on their own data quality and business requirements.
Industry trends: Inference efficiency competition intensifies, open-source ecosystem continues to grow
Looking at the AI industry dynamics in September 2026, inference efficiency has become a core metric for vendors to compete on. From the cloud to the edge, major models highlight tokens generated per second, time to first token, and memory usage. Behind this trend are users' rigid demand for real-time interactive experiences and the strict constraints on resource consumption in edge computing scenarios. Industry experts point out that future competition in model architecture will shift from parameter count to "intelligence output per unit computing power," that is, how to achieve stronger inference capabilities under limited computing power.
The open-source ecosystem also saw significant progress in September. Multiple open-source large model projects released new versions supporting multimodality and tool calling, and the number of community-contributed adapters and fine-tuning scripts grew rapidly. Some companies began combining open-source models with commercial products, forming an "open-source base + commercial service" model, which not only reduces customer trial costs but also generates revenue through value-added services.
Overall, the AI industry in September 2026 exhibited characteristics of "lightweight, tool-oriented, and scenario-specific." The maturity of on-device models brings AI closer to users, the refinement of agent frameworks makes AI evolve from "able to talk" to "able to act," and the proliferation of low-code tools accelerates the release of technological dividends. Although the ceiling of large model technology has not yet been reached, exploration in engineering deployment and business closed loops has entered a critical stage.