Schedule: September 13

This presentation provides a comprehensive overview of Ascend chip evolution: from ensemble approach to swarm tactics, and CANN evolution journey from manual investment to intelligent automation. We explore the transition from usable to user-friendly: CANN feature introduction and open-source strategy, including detailed technical roadmap, community engagement plans, and the strategic vision for making CANN a cornerstone of the open AI compute ecosystem. The talk covers architectural innovations, developer experience improvements, and collaborative opportunities for industry partners.

Welcome Speech (20 min): Prof. Tiejun Huang (Professor from Peking University, Chairman of BAAI)
Opening Keynote (20 min): Dr. Mehdi Snene (Chief AI Officer at UN ODET)

The opening session sets the stage for a day of exploration into how open technologies can drive sustainable development globally.

Host Opening

OpenSeek aims to unite global open-source communities to collaboratively advance algorithms, data, and systems for next-generation models. In stage one, we released CCI4.0, OpenSeek-small, and OpenSeek-Pipeline, developed the DualPipe-V strategy in FlagScale, and launched a collaborative task with the PAZHOU Competition to support community contributions. The initiative is organized into three dedicated working groups, focusing respectively on data, algorithms, and systems.

In a world where massive AI projects are closed-source, and a handful of companies are set to control the future of AI, people are rightfully asking, "Where is OpenCV 5?" Join us to learn about the past, present, and future of the open-source computer vision library we all use and love. We will discuss what is new and improved in OpenCV 5 and what is deprecated. Finally, we will discuss OpenCV's future roadmap.

W3C works with global Web community on analyzing the broad and evolving impact of AI systems on the Web, and examining the potential role of Web standardization in shaping and managing that impact. This talk shares the recent AI related activities and discussions in W3C such as AI in browser, AI Agents, Web ML, AI and accessibility, AI security and privacy, as well as some coming activities on the topic of Web and AI.

Agents are facing a year of comprehensive explosion. This presentation will focus on the full lifecycle of Agents, exploring how to better help Agent developers quickly take Agents from idea to demo, and from demo to production, covering development, evaluation, observation, optimization, and enterprise-level applications.

The advent of the age of artificial general intelligence has changed the landscape and paradigm of education. What are the new challenges that education faces in the AI era? Where are the possible solutions? The Super Brain AI Incubator has conducted a series of educational explorations, discussing the fundamental changes in education in the AI era and validating various possibilities for education.

Opening Remarks

Flutter is one of the most exciting cross-platform frameworks today—but anyone who has worked closely with it knows the pain of waiting for upstream fixes, struggling with forks, and maintaining patches. In this session, we'll explore Nest: a toolkit for extending Flutter's engine and framework without fracturing the ecosystem. We'll share why Nest exists, how it works, and the challenges we've faced in building it, including technical design, CI/CD pipelines, and the practical realities of hosting engine binaries for developers who need faster iteration.

A Secure and Sustainable Future

Servo is a new web rendering engine written in Rust. This workshop will be an in-depth look at Servo's one-of-a-kind layout system, which uses Rust's fearless concurrency to add parallelism to layout. We'll explore Servo's layout engine to understand how it works, driving deep into the technical details, explaining the architecture of a modern CSS layout engine, looking at how CSS features are implemented, and describing the different phases and data structures involved.

As the scale of large language models and their training data continues to expand, the bottleneck between model performance and resource consumption is becoming increasingly prominent. Traditional approaches to model optimization often focus on architectural improvements or hardware acceleration, but overlook the fundamental question of how knowledge is distributed and utilized within these massive parameter spaces. This presentation introduces a novel computing framework driven by knowledge density analysis, which enables more efficient utilization of computational resources by identifying and prioritizing the most knowledge-rich components of large language models during both training and inference phases.

Open Atom Open Source Xuanwu Community's Ecological Vision

Servo is a web engine without a browser. In this workshop we will try to remedy this by embedding Servo in various browser-like applications, and showcase various potential use cases (maybe even the AI browser from the future).

Panel Discussion (60 min)
Moderator: Yonghua Lin (VP and Chief Engineer at BAAI)
Panelists:

• Nicolas Flores Herr (Team Lead Foundation Models & Gen AI Systems, Fraunhofer IAIS)
• Joaquin Salvachua (Professor at Universidad Politécnica de Madrid, ML-OPS & Big Data expert)
• Johann Diedrick (Mozilla Leading Scientist, Open Data Initiative at Mozilla)
• Richard Bian (Director of Strategy and New Product (OSS), Ant Group)
• Yong Qin (Prof., Nankai University)

Exploring how open data initiatives can accelerate progress toward the UN SDGs through improved data sharing and collaboration.

This is a deep-dive for the builders, architects, and infrastructure experts. We will take the audience on a technical journey into the heart of the Bielik.AI project. We start with the "engine": our custom PyTorch framework built specifically for supercomputer environments. This presentation explores the architectural decisions, optimization strategies, and engineering challenges involved in adapting PyTorch for large-scale distributed training on supercomputing infrastructure. We examine performance bottlenecks, memory management strategies, and the practical considerations of scaling deep learning workloads to thousands of compute nodes.

Based on the foundation of globalized production, supply chain, and sales networks built from electric off-road vehicles, this talk describes how Kandi Technologies, now serving as an integrator in the intelligent robotics field, anchors North American security patrol and golf caddy robots. It explores the journey from manufacturing overseas to intelligent manufacturing overseas, and then to ecological overseas expansion, becoming China's technology bridge.

With the development of large model technology, agents are gradually becoming core participants in the Internet. This presentation reviews the technical evolution from the Semantic Web to the Agentic Web, proposing the urgency and necessity of building standardized agent network protocols. We identify four key trends of the Agentic Web: agents replacing traditional software, achieving universal interconnection, protocol-centered connection methods, and autonomous collaborative capabilities. We propose design principles and functional requirements for agent network protocols, and compare mainstream protocols like MCP, A2A, ACP and ANP.

In the current era of rapid iteration of large model applications, large model services represented by MaaS are increasingly becoming one of the core ways for developers to call large model capabilities for application development due to their convenient access methods and low usage thresholds. However, facing the emerging and distinctive large model service platforms in the market, developers often face significant selection difficulties. How to find stable and cost-effective service providers has become an important topic. This presentation shares findings based on market demand and the difficulties developers face. Tsinghua University, in collaboration with the China Software Evaluation Center, launched the '2025 Large Model Service Performance Rankings' to provide comprehensive evaluation and guidance for the industry.

Exploring the current limitations of AI painting, what are the fundamental differences between AI and human thinking, and what directions the future may hold.

Rust x AI: Creating Safe and Efficient Next-Generation Strategy Tools for Small and Medium Investors

Web automation uses intelligent agents to perform high-level tasks by mimicking human interactions with webpages. Despite recent advances in LLM-based web agents, efficiently navigating complex, real-world webpages remains challenging due to massive DOM structures. This talk will present Prune4Web, a paradigm that transforms DOM processing from LLM-based filtering to programmatic pruning. This approach eliminates the need for LLMs to process full DOMs, instead delegating traversal and scoring to lightweight, interpretable programs.

The explosive growth of artificial intelligence applications urgently requires improving development productivity while ensuring performance. This presentation introduces AsNumpy, a lightweight Python mathematical computing library jointly developed by Harbin Institute of Technology and Huawei that deeply supports Ascend NPU and is highly compatible with the NumPy interface. The presentation first analyzes the current computing modes of artificial intelligence, then presents the design scheme and characteristics of AsNumpy, and finally demonstrates the ease of use of AsNumpy through examples.

Vision Doc: The Next 10 Years of Rust

Servo is ported to OpenHarmony in 2024, building upon the progress presented at GOSIM 2024. This talk demonstrates the unified rendering approach of Servo on OpenHarmony and showcases significant advances made for WebDriver integration with automated testing capabilities. We explore the technical architecture of unified rendering across different device form factors, the implementation of WebDriver standards for cross-platform automation, and the roadmap ahead for enhanced web compatibility and performance optimizations in the OpenHarmony ecosystem.

This presentation will reveal the key findings of the Khronos Group's comprehensive AI Ecosystem Research Project, a major initiative undertaken to gain a data-driven understanding of the rapidly evolving AI acceleration landscape. We explore how Khronos standards like OpenCL, Vulkan, and emerging specifications address the diverse needs of AI developers across different hardware platforms. The presentation covers industry trends, hardware vendor perspectives, developer requirements, and the strategic role of open standards in ensuring interoperability and innovation in AI acceleration technologies.

This talk presents Amiga, a next-generation rover platform, specifically designed for modular robotics in farming and outdoor logistics. Engineered for rapid prototyping and real-world deployment, Amiga enables customizable electric utility vehicles that streamline labor-intensive tasks, reduce costs, and support organic regenerative farming practices. We explore Amiga's hardware and software modularity, highlighting interchangeable toolkits and actuation systems, and demonstrate how the SDK offers easy-to-use APIs and edge-ML pipelines that power agentic AI and visual-language action models for dynamic environmental responsiveness.

Servo is a modern open source web rendering engine written in Rust. Built for safety, modularity, and performance, Servo brings memory safety and concurrency to the forefront of browser engine innovation. Servo development has been reactivated in 2023 since Igalia took over the maintenance of the project, the project community has been constantly growing since then. In this talk we'll discuss the last few years working on the project, the main features implemented and the overall evolution of Servo. We'll highlight the new platform additions like Android and OpenHarmony, together with the multiple optimizations that have been developed to make Servo a viable alternative to other web engines.

Cognitive AI represents a transformative leap in how machines understand and interact with the world. Despite its potential, practical challenges remain in making these systems accessible and applicable across diverse domains. This talk addresses how multimodal models, combined with Retrieval-Augmented Generation (RAG) and agentic workflows, can enable cognitive AI systems to deliver personalized, context-aware solutions with OpenVINO. With applications ranging from educational tools to assistive technologies for the elderly and disabled, this talk focuses on practical strategies for optimizing, and deploying these models and pipelines, making them both scalable and accessible to researchers and practitioners.

Bridging Art and Artificial Intelligence : Augmenting Creativity ?

This presentation mainly introduces the practical implementation of iFLYTEK's Spark large model using Ascend computing power. It focuses on reporting how Spark's large-scale training clusters achieve rapid system recovery under ten-thousand-card Ascend clusters, and the development of efficient fusion operators using Ascend C technology.

W3C works with the global Web community on analyzing the broad and evolving impact of AI systems on the Web, and examining the potential role of Web standardization in shaping and managing that impact. This talk explores how AI is transforming web technologies, the emergence of intelligent web agents, advances in Web Machine Learning APIs, accessibility considerations for AI-powered web applications, and security and privacy challenges in the AI-web intersection. We'll discuss upcoming W3C initiatives and standardization efforts that will define the future of intelligent web experiences.

Institutional memory and continuity through Rust specifications

Show & Tell #1 (10 min)
Bangxu Yu (SVP of CSDN & CEO of GitCode)

Show & Tell #2 (10 min)
Xinwei Hu (Chair of the openEuler Community Technical Committee)

Exploring how open-source foundational software can enable collaborative ecosystem building for AI development.

Welcome Speech

Chitu inference engine is a large model inference engine that supports multi-element computing power and is actively developed in the open source community. This report introduces the birth history, technical evolution and future vision of the Chitu inference engine.

Panel Discussion (60 min)
Moderator: Anni Lai (Co-Chair, Generative AI Commons)
Panelists:

• Nooman Fehri (Former Digital and Technology Minister, Tunisia)
• Vincent Caldeira (CTO of Red Hat APAC, Top 10 CTO in APAC 2023)
• Matt White (Executive Director, PyTorch Foundation)
• Richard Lin (Co-founder of Kaiyuanshe, Head of the RTE Developer Community)
• Liyun Yang (Mulan Community - Open Source Policy Strategy for China Government)

Discussing governance models, contribution pathways, policy and legal considerations, and alignment mechanisms for UN engagement with open source communities.

This presentation will discuss the models open-sourced by ZhiPuAI since 2025, including the GLM-4.5 series and GLM-4.1V series, as well as describe ZhiPuAI's engineering and algorithmic adaptation in model open-source, and community collaboration models. Through a systematic explanation of ZhiPuAI's open-source model processes, developers will become familiar with the engineering and algorithmic challenges in the large model open-source process, solutions, and working methods.

Panel ：Embodied Intelligence — The Second Half of AI

As AI technology enters a new era, the limitations of traditional digital architecture become increasingly apparent. This presentation will introduce the revolutionary RVP (Reality-Virtualization-Perception) protocol, a brand-new AI ecosystem architecture designed for the post-digital era. The talk will deeply analyze the three core components of the RVP protocol: multi-source data fusion at the reality layer, digital twin technology at the virtualization layer, and multimodal AI capability integration at the perception layer. It will highlight how this protocol breaks through existing technical bottlenecks to achieve cross-domain collaboration and adaptive evolution of intelligent networks.

SPEAR - A Scalable and Distributed AI Agents Framework Across Edge and Cloud

Interdisciplinary Research on the Integration of Artificial Intelligence and Traditional Art

Overall progress of the Cangjie programming language ecosystem development.

Intro

React Native apps are bigger and more complex than ever. At Callstack, we've built Rock—open, modular, self-hosted, and incrementally adoptable framework for React Native—to help teams scale, migrate, and deliver faster. React Native just turned 10, and its adoption has followed a familiar curve—from early startups to scale-ups, mainstream apps, and finally large enterprises making their move. The Community CLI has long been the backbone of this ecosystem, but as React Native evolved—embracing Expo, modularity, and framework-driven development—we realized the needs of today's teams go far beyond what the CLI alone can offer. In this talk, we'll share how we got here, why frameworks are the future of React Native, and how our new open-source framework can help teams migrate smoothly, unify tooling, and accelerate delivery.

Communication At Any Size

Towards an Easy Rust

Asterinas: A Rust-Based Framekernel to Reimagine Linux in the 2020s

How llama.cpp works, their focus, roadmap, how to contribute, etc

This talk is about parallel inference technologies and optimizations on LLM inference and optimization practices on Chitu inference system.

Ant AI Gateway - A Practical Guide to Optimizing Performance in Large-Scale Online Inference Service Clusters

Aesthetics serve as an implicit and important criterion in song generation tasks that reflect human perception beyond objective metrics. However, evaluating the aesthetics of generated songs remains a fundamental challenge, as the appreciation of music is highly subjective. To address this issue, we introduce SongEval, the first open-source, large-scale benchmark dataset for evaluating the aesthetics of full-length songs. SongEval includes over 2,399 songs in full length, summing up to more than 140 hours, with aesthetic ratings from 16 professional annotators with musical backgrounds. Each song is evaluated across five key dimensions: overall coherence, memorability, naturalness of vocal breathing and phrasing, clarity of song structure, and overall musicality.

MonkeyCode-Open-Source AI code assistant

Cangjie in HarmonyOS applications: Nami AI search application development practices.

The Mixture of Experts (MoE) architecture enables scaling Large Language Models while keeping computation costs low. However, ultra-large MoE models with hundreds of billions of parameters require massive memory and storage, making edge deployment challenging. This presentation introduces a comprehensive compression framework combining expert pruning, MoE-specific mixed-precision quantization, and activation optimization. The framework reduces both model weight size and activation memory usage, achieving the first efficient deployment of models as large as DeepSeek-V3 under 128 GB memory constraints, outperforming uniform low-bit quantization methods.

Crate security in 2025: tools and techniques

High-Level Rust and the Future of Application Development

Sound, Bare Metal, Rust, AI and A Challenge

A data space definition language (DSDL) to enable interoperability between data spaces and enable the integration with the Agentic AI Ecosystem.

Building the Next-Generation AI Infrastructure Device Manageability Based on openUBMC

As gaming continues to evolve toward richer, more immersive experiences—including real-time rendering, volumetric environments, and AI-driven interactivity—efficient compression of both neural models and game data becomes critical. This presentation explores cutting-edge AI-driven compression techniques specifically designed for next-generation gaming applications. We examine how machine learning approaches can optimize asset storage, reduce bandwidth requirements, and maintain visual fidelity while enabling seamless streaming of complex game content across diverse hardware platforms.

Sharing Chitu large model industrial deployment practices: discussing tool-chain efficiency improvement cases and providing reusable implementation experience.

Nexent: How to Build Agents with AI, and How Agents Can Be Applied in Real Enterprises

Cangjie in HarmonyOS: Sharing of iQIYI Application Development

Hugging Face transformer Optimum project

Show & Tell #3 (10 min)
Alex Zhu (Front-end Engineer, Ant Group | Ant Design Contributor)

Empowering Cloud-Native AI: Solving Large-Scale LLM Deployment Challenges with the Volcano Scheduler

Mixing of multi modal AI model to solve complex robotic task with dora-rs

Towards Trustworthy and Reproducible Large Models: An Open-Source Tiered Framework

YiCloud INESA Smart Computing Platform Overview: A training-inference integrated smart computing cloud platform for AI application deployment, dedicated to solving core pain points from AI model development to large-scale deployment, achieving an efficient "development-to-online" closed loop. Features include unified training/fine-tuning/inference support, intelligent heterogeneous resource scheduling (NVIDIA/Ascend/MeTaX chips), elastic inference services, and full-stack domestic adaptation.

The Fastest Legend: Developing the Global Cloud Database ScopeDB with Rust in Four Months

A WebAssembly Ahead-Of-Time On-Target Translator for Aurix Microcontrollers

Application and Open Source of Auron, a Big Data Vectorization Engine Implemented with Rust

Cangjie in University: Practical Teaching of Open Source Software Engineering

Panel Discussion (60 min)
Moderator: Alexy Khrabrov (AI/ML Community Architect, Neo4j)
Panelists:

• Dr. Minghui Zhou (Tenured Full Professor and Vice Dean, School of Computer Science, Peking University)
• Mohamed Farahat (Former UNSG HLAB member, Legal Expert with UNESCO)
• Walid Mathlouthi (Head of Infrastructure Division, ITU)
• Xinwei Hu (Chair of the openEuler Community Technical Committee, Huawei)
• Yonghua Lin (VP and Chief Engineer at BAAI)

Exploring how Open Science principles can accelerate progress toward UN SDGs through transparency, collaboration, and accessibility. Focus on barriers, enablers, cross-border collaboration, and ethical considerations.

verl: An Open-Source LLM Reinforcement Learning Framework for Agent-Oriented Training

From Investigation to Knowledge: A New Paradigm for Embodied Intelligent Robot Development

Unlock the infinite potential of AI agents! This keynote presentation will dive deep into how to leverage ADK, A2A, MCP and Agent Engine to build AI agents. Learn how to harness Google's latest AI technologies to create collaborative, efficient, and scalable multi-agent systems. Explore the future of agent development and discover how agents will revolutionize the way we interact with technology. This comprehensive overview covers Google's strategic vision for the agentic web, practical implementation frameworks, and the collaborative ecosystem needed to realize the full potential of AI-powered web interactions.

At a time when industry intelligent needs are upgrading and device barriers need to be broken, this sharing focuses on HarmonyOS industry innovation solutions: starting from its distributed technology and full-scenario collaborative capabilities, it explains HarmonyOS app industry solutions in detail to help enterprises reduce development costs and increase efficiency.

Generative AI in education : Open source solutions and challenges.

KTransformers is a CPU-GPU heterogeneous inference framework that can perform inference on mainstream large models such as DeepSeekR1 and KimiK2 using a single card. It achieves separation of different computations by placing MoE layers on CPU and MLA on GPU, fully utilizing resources of different hardware. Additionally, KTransformers adopts the newly developed Expert Defer technology, which can fully leverage the advantages of CPU-GPU heterogeneous architecture and significantly improve performance. KTransformers has made extensive attempts on different hardware platforms and achieved good results.

Deployment and Operational Practices for Chitu's High-Performance PD Disaggregation Architecture Inference Service Based on K8S

Rewriting FreeSWITCH with Rust

Deploying Rust UIs to the Terminal（and your pocket）

Makepad Rust applications and AI

Spire Tianqing, as the first Web development framework designed specifically for enterprise-level applications, provides Web development technology solutions for developers by integrating modern Web development concepts with Cangjie language features. In terms of architectural design, Tianqing's overall design borrows from ASP.NET Core's modular design philosophy, adopts modern patterns such as middleware pipelines and dependency injection, and achieves excellent performance and horizontal scalability while maintaining a lightweight kernel. The framework significantly reduces the threshold for Cangjie language developers to build complex Web applications through carefully designed API interface layers.

With the vision of 'making machines think like humans and making human life better with trustworthy AI,' ZhiPuAI is committed to building a new generation of cognitive intelligence large models, focusing on Chinese innovation in large models. This talk covers GLM's evolution from imagination to productivity, including GLM-Realtime, AutoGLM, and GLM-4.5 achievements.

Leveraging Rust and Khronos' emerging Slang initiative, we introduce our initial efforts toward a cross-platform GPU LLM inference ecosystem. With a single-source approach we aim to minimize backend-specific code and foster community participation.

Solana: Built for Rustaceans

Exploring a conceptual new incremental system for the Rust compiler

Developing BlueOS with Rust and Open Source Practices

High-acceptance-rate speculative inference schemes represented by EAGLE/MTP are driving the deployment of speculative inference. Speculative inference computes multiple tokens in one inference, which can fully leverage Ascend's high computational density-to-bandwidth ratio. To this end, we developed the high-performance inference framework omniinfer to fully unleash Ascend's performance. Targeting the model structural characteristics of speculative inference schemes like EAGLE and MTP, we optimized the speculative inference scheduling framework, reducing Ascend's idle time, and optimized sampling methods to maintain model accuracy while improving acceptance rates. We also implemented hardware-specific optimizations for Ascend's characteristics.

Human-Agent Interaction in Business: Autonomy vs. Human Control

Conversations Without Boundaries: Building Voice Agents with Flutter, WebRTC, and LiveKit

In this presentation, we present the recent progress of our end-to-end AI chat and translation system on local devices, such as laptop computers or local AI accelerators. Our end-to-end system has innovative features including specialized model architectures for efficient inference on edge devices, advanced optimization techniques for real-time performance, and comprehensive multilingual support with high-quality translation capabilities. We demonstrate practical applications and performance benchmarks that showcase the viability of running sophisticated AI systems entirely on local hardware without cloud dependencies.

Education 4.0: A New Age of Creativity with AI

AI Healthcare Platform Application Development Based on Cangjie

Ascend CANN open source strategy and hierarchical decoupled technology stack sharing, focusing on deepening cooperation in the Ascend technology ecosystem and building competitive advantages in the third-party ecosystem.

Show & Tell #4 (10 min)
Yao Chen (COO, OYMotion)

Edge AI enables real-time, low-latency inference by processing data locally, unlocking transformative applications across industries. With advancements in cloud-native technologies, Edge AI is evolving into a powerful cloud-edge collaborative paradigm, allowing dynamic AI workload orchestration between edge and cloud to optimize performance, accuracy, and privacy.

Opening Talk (10 minutes)

• Title: Artificial Intelligence As a Digital Public Goods
• Speaker: Serge Stinckwich (Head of Research, United Nations University Institute in Macau)

Panel Discussion (50 min)
Moderator: Bryan Che (Board Member, Eclipse Foundation and LF CNCF)
Panelists:

• Prof. Kai Du (Assistant Professor, Psychology & Cognitive Science, Tsinghua University)
• Prof. Yin Peng (Assistant Professor, Department of Data Science, CUHK)
• Dr. Satya Mallick (CEO of OpenCV.org, AI & Computer Vision Expert)
• Adina Yakefu (EA / AI researcher, Hugging Face)
• Serge Stinckwich (Head of Research, United Nations University Institute in Macau)

Envisioning the future of AI in sustainable development context. Focus on inclusive and equitable AI futures, bridging innovation with impact, and promoting multistakeholder collaboration.