MediaTek & Alibaba Cloud Deploy AI Model on Tianji Platform

March 28, 2024 – MediaTek has announced an in-depth collaboration with Alibaba Cloud, achieving the milestone deployment of a smaller-scale model of Alibaba’s Tongyi Qianwen (aka “Comprehensive Thousand Questions” AI model) on MediaTek’s Dimensity 9300 mobile platform. This deployment is compatible with the Dimensity 8300 platform and facilitates instantaneous, accurate multi-turn human-machine dialogue in offline mode. Looking ahead, both parties will partner to build a generative AI software and hardware ecosystem targeting app developers and device manufacturers, based on the MediaTek Dimensity mobile platform adapting more parameter versions of Tongyi’s large models, exploring new opportunities for AI Agent services accessible to the general public.

Dr. Li Yanji, Vice President of MediaTek’s Wireless Communications Business Unit, stated: “The Tongyi series of large models from Alibaba Cloud are AI frontrunners. We anticipate that our collaboration will provide app developers and terminal customers with more potent hardware and software solutions. This will also expedite the deployment of generative AI on devices, as well as the rapid growth of AI applications and AI Agent ecosystems, delivering exciting new AI product experiences to users.”

Xu Dong, the person in charge of Alibaba’s Tongyi Laboratory, mentioned: “On-device AI is one crucial scenario for the deployment of large models, but it faces numerous challenges such as difficulties in software-hardware compatibility and incomplete development environments. In our partnership with MediaTek, we’ve surmounted various technical and engineering challenges at the foundational and development levels, effectively embedding large models into smartphone chips and setting a successful example for Model-on-Chip deployment of on-device AI in the industry. Based on this, we believe developers will be inspired to create even more innovative and practical AI applications.”

Pioneering the on-device deployment of generative AI by joining forces in software and hardware optimization for enhanced efficiency

In contrast to applications and services that deploy generative AI only in the cloud, hybrid edge-cloud AI can fully leverage the computing power of end devices, offering distinct advantages in terms of operating costs, user information security, real-time performance, and personalized user experiences. With insightful industry trends, MediaTek and Alibaba Cloud’s Tongyi Lab have pursued an in-depth cooperation to adapt and optimize the Tongyi large models to run on MediaTek’s Dimensity platform-end devices, empowering these devices to provide safer, faster, reliable, and differentiated user experiences.

The Dimensity 9300 integrates MediaTek’s seventh-generation AI processor APU 790, which features a generative AI engine at the hardware level, supporting the running of AI large language models from 1 billion, 7 billion, and 13 billion up to 33 billion parameters on the device. It also supports the expansion of generative AI model skills with the NeuroPilot Fusion technology. The processing speed of generative AI on APU 790 is eight times faster than its predecessor processors, capable of generating images within one second. Furthermore, MediaTek’s NeuroPilot AI development platform has established a rich AI ecosystem, compatible with the world’s mainstream large AI models, and its complete toolchain offers support for the NeuroPilot Fusion and LoRA (Low-Rank Adaptive fusion) technologies. Combined with the strong computing power of MediaTek’s AI system, this empowers developers to quickly optimize and transform models, significantly improving the on-device processing efficiency of large models while endowing them with a more comprehensive capacity.

Tongyi Qianwen is a self-developed LLM foundational model by Alibaba Cloud, which has released a 100-trillion-parameter version 2.0 and open-sourced versions with 72, 14, 7, 4, 1.8, and 0.5 trillion parameters, along with multimodal large models such as Qwen-VL for visual understanding and Qwen-Audio. The Qwen series of Tongyi Qianwen models are among the first large models open-sourced in China, achieving outstanding performances in over ten industry benchmark evaluations and topping the global Hugging Face rankings in terms of impact. The Tongyi model family has formed a matrix of foundational models and industry-specific models, and it was among the first to pass the national large model evaluation standards. In all 11 tests of professional AI large model reports, it scored full marks in six items, ranking among the top.

MediaTek has maintained a close partnership with Alibaba Cloud for an extended period, and together, they have accomplished the on-device adaptation of the small-scale version of the Tongyi Qianwen model on the Dimensity mobile platform, soon to be promoted across on-device AI applications via the developer ecosystem. Additionally, MediaTek’s developer-oriented model centers and toolchains will integrate the Tongyi Qianwen model series to facilitate developers in more conveniently developing a wide range of scenario applications based on the Tongyi large models on the Dimensity mobile platform.

MediaTek and Alibaba Cloud’s pioneering alliance with the Tongyi Qianwen large model explores new opportunities for AI Agent applications.

AI Agents, the most thrilling evolutionary direction for generative large models – “the next destination for large models” – are a key focus of this collaboration. With a shared understanding of new opportunities, both parties have announced the initiation of a joint exploration program for AI Agent solutions. By integrating the powerful hardware and AI computational capabilities of MediaTek’s Dimensity mobile platform with the advanced large model technology and ever-improving development tools platform from Alibaba Cloud’s Tongyi Qianwen, they aim to support intelligent terminals in achieving high-efficiency on-device natural language understanding, complex decision-making, and personalized service generation. The goal is to explore the creation of the next-generation intelligent terminal application ecosystem with context awareness, autonomous learning, and real-time interaction functions.

In the future, efforts by MediaTek and Alibaba Cloud will allow users to voice their needs at home and have the AI Agent, powered by the Dimensity mobile platform and Tongyi large model, autonomously establish and ultimately accomplish tasks, turning the dream of intelligent living into reality. Furthermore, this will open up broader avenues for innovation for developers and terminal device manufacturers.

Amidst the surging wave of generative AI, large models are becoming a vital driving force for a new technology revolution and industrial transformation. MediaTek will continue to join hands with Alibaba Cloud’s Tongyi Lab to accelerate the deployment of generative AI on the endpoint, nourish application scenarios and ecosystems for generative AI, and co-create future-oriented AI Agent product experiences. This partnership aims to support the widespread application of AI across numerous industries, enabling technology to benefit everyday life.

Leave a Reply

Your email address will not be published. Required fields are marked *

This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.