MediaTek and Alibaba Cloud have completed the device-side deployment of the Tongyi Qianwen model on the Dimensity mobile platform

On March 28, 2024, MediaTek announced that it has reached an in-depth cooperation with Alibaba Cloud to complete the device-side deployment of the small-sized version of the Tongyi Qianwen large model on the Dimensity 9300 mobile platform, which can be adapted to the Dimensity 8300 mobile platform to enable real-time and accurate multi-round human-machine dialogue Q&A in offline mode. In the future, the two parties will work together to build a generative AI software and hardware ecosystem for application developers and device manufacturers, and jointly explore new opportunities for AI agent service scenarios based on the MediaTek Dimensity mobile platform with more parameter versions of the Tongyi large model.

"We look forward to working together to provide more powerful hardware and software solutions to application developers and end customers, as well as facilitate the deployment of generative AI on the device side, as well as the rapid development of AI applications and AI agent ecosystems, bringing more exciting AI product experiences to users," said Dr. Yanji Li, Vice President of MediaTek's Wireless Communications Business Unit. ”

Xu Dong, head of Alibaba Tongyi Lab, said: "End-to-device AI is one of the important scenarios for the application of large models, but it faces many challenges such as difficult adaptation of software and hardware, and incomplete development environment. Alibaba Cloud and MediaTek have worked together to overcome a large number of technical and engineering problems in low-level adaptation and upper-layer development, truly 'loading' large models into mobile phone chips, and successfully proofing the new model-on-chip deployment model of end-side AI for the industry. Based on this, we believe that developers will create more novel and practical AI applications in the future. ”

Proofing generative AI device-side deployment, software and hardware integration optimization to improve efficiency

Compared with applications and services that only deploy generative AI in the cloud, device-cloud hybrid AI can make full use of the computing power of terminal devices, and has significant advantages in terms of operating costs, user information security, real-time, and personalized user experience. Based on consistent insights into industry trends, MediaTek and Alibaba Cloud Tongyi Lab have collaborated to adapt and optimize the Tongyi model running on Dimensity mobile platform devices, enabling devices to provide a safer, faster, more reliable, and more differentiated user experience.

The Dimensity 9300 integrates MediaTek's 7th generation AI processor APU 790, which has a built-in hardware-level generative AI engine that supports devices to run AI large language models with 1 billion, 7 billion, 13 billion, and up to 33 billion parameters, and supports NeuroPilot Fusion, the "skill enrichment" technology on the device side of generative AI models, which is 8 times faster than the previous generation of AI processors and can generate images in 1 second. In addition, MediaTek's AI development platform, NeuroPilot, has built a rich AI ecosystem, which has been adapted to the world's mainstream AI models, and the complete toolchain includes support for NeuroPilot Fusion and LoRA (Low-rank Adaptive) fusion technology, combined with the powerful computing power of MediaTek's AI system, to help developers complete model optimization and transformation faster, significantly improve the device-side processing efficiency of large models, and empower them with more comprehensive capabilities.

Tongyi Qianwen is an LLM base model developed by Alibaba Cloud, which has launched version 2.0 with 100 billion parameters, 72 billion, 14 billion, 7 billion, 4 billion, 1.8 billion, and 500 million parameters of open source, as well as multi-modal large models such as visual understanding model Qwen-VL and audio large model Qwen-Audio. Tongyi Qianwen Qwen series is the first open-source self-developed large model in the Chinese market, and has achieved outstanding results in more than 10 industry benchmark evaluations, and topped the global high-impact Hugging Face rankings. At present, the Tongyi large model family has formed a matrix of basic large models and a series of industry models, and has passed the national large model standard evaluation for the first time, and has obtained 6 full scores in all 11 tests of the industry's professional AI large model report, ranking among the best.

MediaTek and Alibaba Cloud have maintained a close partnership for a long time, and the two parties have completed the adaptation of the small-size version of the Tongyi Qianwen large model to the Dimensity mobile platform, and will promote the widespread popularization of this achievement in device AI applications through the developer ecosystem. In addition, MediaTek's model center and toolchain for developers will integrate the Tongyi Qianwen model series, making it easier for developers to develop applications based on the Tongyi model on the Dimensity mobile platform.

MediaTek has partnered with Alibaba Cloud Tongyi Qianwen Model to pioneer new opportunities for AI agents on the device side

AI agents are the most exciting development of generative large models, and are known as the "next stop of large models". Based on the unanimous understanding of the new opportunities, the two parties also announced the launch of a joint exploration of AI agent solutions, which integrates the powerful hardware and AI computing capabilities of MediaTek's Dimensity mobile platform, as well as Alibaba Cloud's advanced large-scale model technology and increasingly complete development tool platform, to support intelligent terminals to achieve natural language understanding, complex decision-making, and personalized service generation in an energy-efficient manner on the device side. The next-generation intelligent terminal application ecosystem with self-learning and real-time interaction functions.

In the future, through the joint efforts of MediaTek and Alibaba Cloud, users can "move their mouths" to express their needs at home, and AI agents driven by the Dimensity mobile platform and Tongyi large model will be able to automatically build and finally achieve tasks, turning the dream of intelligence into reality. At the same time, developers and terminal equipment manufacturers will also usher in broader innovation opportunities.

Under the wave of generative AI, large models are becoming an important driving force for a new round of technological revolution and industrial transformation. MediaTek will continue to work with Alibaba Cloud Tongyi Lab to accelerate the deployment of generative AI on the device side, jointly cultivate generative AI application scenarios and ecosystems, and jointly create future-oriented AI agent product experiences, so as to better support AI applications in thousands of industries and bring technology to people's lives. (Xuan Hongyan)