CNCC 2026: on-device model forum opens tomorrow to explore new frontiers of on-device intelligence

2026-10-21

About the forum

Since ChatGPT launched in late 2022, large model technology has advanced rapidly and model scale has kept expanding. AI's deep involvement in R&D work such as model architecture, training frameworks and data synthesis has also pushed large models toward higher capability density. As their capabilities grow, powerful intelligence is moving from the cloud to the edge at an accelerating pace. On-device large models are becoming an important technical foundation for intelligent terminals, evolving from "able to converse" toward on-device intelligence that "can perceive, reason and act."

Yet real-world deployment of on-device large models still faces many challenges. On one hand, edge devices are constrained by compute, storage and power, and must achieve higher model capability density and more efficient training and inference within limited resources. On the other hand, existing large models still fall short in real-time perception, multimodal interaction and autonomous action. At the same time, the architecture and resource constraints of different edge chips vary widely, making efficient coordination between models and hardware-software systems a major issue. How can we break through compute, storage and power constraints to achieve efficient training and inference? How do we move from "able to converse" to real-time omni-modal interaction, autonomous perception, reasoning and action? This forum focuses on the key technologies and deployment practices of on-device large models, exploring frontier directions such as device-cloud collaboration, agents, chip-level customization and speech generation, and discussing the trends and future challenges of on-device intelligence. The forum will examine the key technologies and development directions of on-device large models, with a focus on autonomous agents, real-time omni-modal interaction, building high-capability-density models, chip-level training and inference customization, and speech generation, while exploring the technology roadmap and application prospects of on-device intelligence.

Agenda: Time: October 22, 2026, 13:30-17:30 · Venue: Hall 5-502, Western China International Expo City, Chengdu

Scan the QR code below to watch the livestream:

Forum Chair

/ Liu Zhiyuan /

Liu Zhiyuan is a CCF Distinguished Member and a tenured professor in the Department of Computer Science and Technology at Tsinghua University. His main research areas are large models, natural language processing and knowledge graphs. He has published more than 300 papers in venues including Nature sub-journals, ACL and EMNLP, with over 100,000 Google Scholar citations. His honors include the First Prize in Natural Science from the Ministry of Education, the First Prize of the Qian Weichang Chinese Information Processing Science and Technology Award from the Chinese Information Processing Society of China, the World Internet Leading Scientific and Technological Achievement Award, and the Science and Technology Progress First Prize of the Wu Wenjun Artificial Intelligence Science and Technology Award. He has been named one of the Eighth Beijing Young Teaching Masters, an Elsevier Highly Cited Chinese Researcher for six consecutive years from 2020 to 2025, a member of MIT Technology Review's 35 Innovators Under 35 China list, and a recipient of the Young Elite Scientists Sponsorship Program by CAST.

Co-Chair

/ Yao Yuan /

Yao Yuan is an assistant professor at the Institute for AI Industry Research, Tsinghua University, focusing on multimodal large models. He has published more than 30 papers at top conferences and journals including Nature Communications, CVPR, ACL and ICML, with over 10,000 Google Scholar citations. He led the construction of efficient multimodal models such as MiniCPM-V and MiniCPM-o, which ranked first on HuggingFace Trending, GitHub Trending and Papers With Code Trending Research for multiple consecutive days. His open-source projects have earned over 39,000 stars in total, and the models have been downloaded more than 34 million times. His honors include the Outstanding Doctoral Dissertation Award of the Wu Wenjun Artificial Intelligence Science and Technology Award and the Intel China Academic Achievement Award.

Forum Speakers

/ Yao Yuan /

Talk title: Toward real-time omni-modal large models with human-like interaction

Abstract: Current multimodal large models are mainly built for offline data understanding and turn-based interaction, and struggle to meet the demands of real-time dynamic interaction in the physical world. Real-time omni-modal large models have therefore become an important path toward general multimodal AI. This talk introduces MiniCPM-o 4.5, which has human-like interaction capabilities: it can see, listen and speak simultaneously during real-time conversation, and can proactively raise reminders and dynamically describe scenes based on continuous scene understanding. Thanks to efficient architecture design and system-level optimization, the model runs smoothly on edge devices. The talk covers model architecture, runtime mechanisms and training methods, discusses key challenges and core design choices, and looks ahead to future trends and application prospects.

/ Xiao Zhaojun /

Xiao Zhaojun is an assistant researcher in the Department of Computer Science and Technology at Tsinghua University, researching efficient large model architectures. He has published multiple papers in top AI journals and conferences including Nature Machine Intelligence, ICML, ICLR, NeurIPS and ACL. He is also responsible for the MiniCPM series of open-source text models and the UltraData series of open-source data, whose models and datasets have been downloaded millions of times and have had broad academic and industrial impact. His honors include the First Prize of the Qian Weichang Chinese Information Processing Science and Technology Award, the Second Prize of the Scientific Research Outstanding Achievement Award for Higher Education Institutions from the Ministry of Education, the Postdoctoral Innovative Talent Support Program, and the Potential Qingyuan Scholar award from the Chinese Association for Artificial Intelligence.

Talk title: Efficient construction of high-capability-density large models

Abstract: The expansion of large language model scale has made high compute and deployment costs a bottleneck for bringing on-device intelligence into practice. Achieving high capability density within limited parameter scale and compute constraints has become an important direction in large model research. Taking the MiniCPM on-device large models as an example, this talk presents a technical system for efficiently producing "high-capability-density large models." Centered on three core questions, model architecture, data governance and learning methods, it discusses compute-efficient architecture design, high-knowledge-density data construction and methods that promote the efficient emergence of intelligence. Drawing on technical solutions and practical experience, it looks ahead to a path toward efficient, accessible on-device large models.

/ Li Yuxuan /

Li Yuxuan is a technical partner at ModelBest and Chief Scientist of Intelligence Evolution. His main research areas are intelligent computing systems, high-performance computing, quantum computing and climate modeling. He has published more than ten papers in high-level international conferences and journals in related fields, with over 4,000 Google Scholar citations. He was one of the lead technical leads for the Sunway TaihuLight supercomputer and has won the world supercomputing competition championship multiple times. His honors include national-level awards such as the China Youth May Fourth Medal (collective).

Talk title: AI making AI: Forge Engineering drives chip-level training and inference customization for on-device large models

Abstract: "AI making AI" is moving from vision to reality, as the software stacks that large models depend on are beginning to be written by AI itself. This matters especially for edge devices, where chips are diverse and constraints vary widely: adapting hardware-software stacks for every chip by hand is extremely costly, and only AI-driven forging can make chip-level customization feasible. This talk introduces the engineering paradigm of AI making AI, Forge Engineering, and uses ForgeTrain, the first production-grade training framework written entirely by AI that surpasses human-written work, to show how months of framework development can be compressed into days. The talk covers forging methodology, correctness assurance and the mechanisms behind performance gains, and looks ahead to paths that accelerate the evolution of on-device intelligence.

/ Zeng Guoyang /

Zeng Guoyang won a CCF NOI gold medal and graduated from the Department of Computer Science and Technology at Tsinghua University in 2020. In 2022, he co-founded ModelBest and serves as its CTO, working to bring large model technology into real applications. He joined the Tsinghua Natural Language Processing Lab in 2017 to work on natural language processing research, and in 2020 joined the "Wudao·Wenyuan" Chinese pre-trained large model team as a core member, taking part in the training and development of the CPM series of models. He wrote the efficient large model training and inference tools BMTrain and BMInf, and holds multiple invention patents. In 2021, he served as deputy director of the Language Model Acceleration Technology Innovation Center at the Beijing Academy of Artificial Intelligence.

Talk title: VoxCPM: a continuous autoregressive speech generation model

Abstract: Speech synthesis is an important technology for on-device intelligence and a bridge and medium for human-machine interaction. In recent years, speech synthesis has been moving from "reading clearly" toward "sounding real." Achieving truly natural, controllable and high-fidelity speech generation across languages and voices has become key to natural human-machine interaction. This talk focuses on the open-source VoxCPM series and introduces a technical route different from mainstream "discrete speech tokens": tokenizer-free speech synthesis. It covers the VoxCPM architecture, actual results and open problems in frontier research, and discusses the future direction of the VoxCPM series and speech generation models.

/ Lin Yankai /

Lin Yankai is a member of the CCF NOI Scientific Committee and a tenure-track associate professor at the Gaoling School of Artificial Intelligence, Renmin University of China. His main research direction is large model agents. He has over 30,000 Google Scholar citations and was named an Elsevier Highly Cited Chinese Researcher and included in Stanford University's annual list of the world's top 2% scientists for six consecutive years from 2020 to 2025. His research has received the First Prize in Natural Science from the Ministry of Education and the World Internet Leading Scientific and Technological Achievement Award, and he has been selected for the Beijing Science and Technology Rising Star program, the Wu Wenjun Artificial Intelligence Youth Science and Technology Award, and the "AI 100 Young Pioneers" list jointly released by MIT Technology Review China and DeepTech.

Talk title: Large model autonomous agents toward general intelligence

Abstract: In recent years, large models have shown enormous potential in understanding and reasoning, moving from a passive "brain" to an active "actor," and building autonomous agents is key to advancing toward general intelligence. Agents need tool use, environment interaction and the ability to execute complex tasks autonomously. This talk focuses on large model autonomous agents and discusses core challenges including generalization, long-horizon planning and autonomy. It reviews the evolution from imitation learning to exploration learning, highlights methods for improving long-horizon planning and environment generalization through reinforcement learning and standardized protocols, and looks ahead to agents developing from "passive tools" into "proactive assistants."

Welcome to CNCC 2026

CNCC is a conference that takes a broad view of technology trends, known for its high profile, large scale and rich content. Its formats include invited keynote reports, plenary forums, special topic forums, featured events and exhibitions. The conference brings together academicians of the Chinese Academy of Sciences and the Chinese Academy of Engineering, leading scholars from China and abroad, and renowned entrepreneurs to look ahead at frontier trends and share innovative results. Representatives of international computing organizations such as ACM, IEEE CS, IPSJ and KIISE have also been invited to attend this event on site many times. CNCC 2026 will be held in Chengdu, Sichuan Province, from October 21 to 24 under the theme "Data Drives, Intelligence Computes the Future." Everyone is welcome to attend.