GOSIM Shenzhen 2026 · Friday (2026-10-16) Detailed Talk Notes

Overview

Dates verified: GOSIM Shenzhen 2026 is held from October 16, 2026 (Friday) to October 17 (Saturday) at the ADEN Hotel Shenzhen Nanshan. It is the seventh stop of the GOSIM series and the first time it is held in Shenzhen. Schedule source: https://shenzhen2026.gosim.org/schedule/; dates and venue source: https://www.bagevent.com/event/gosimshenzhen2026. This document covers only Friday.

Material limitations (important): As of 2026-10-18, no post-event coverage, on-site notes, recordings, or speaker slides are available for any of Friday's talks (searches were conducted in Chinese and English across gosim.org, Zhidx, CSDN, Tencent News, QbitAI, Bilibili, and YouTube; all results were pre-event previews and schedules). Therefore, no entry in this document can be labeled "Coverage/recording available". The three source labels mean:

Day total: 86 talks + 1 Spotlight roadshow = 87 items (deduplicated, excluding lunch and tea breaks).

Track Venue Talks
Keynote Plenary 3F – ADEN (full hall) 3
Agentic AI Summit 3F – ADEN 1 12
Agentic AI on Edge 3F – MING HUA 9
Agentic OS & App 2F – HAI XIA 2 9
Agentic Device 2F – KAI XUAN 1 9
Open Source Models & Infra 3F – ADEN 2 9
Open Source Robotics 2F – HAI XIA 1 10
DORA-rs Workshop 2F – KAI XUAN 2 (morning) 8
CANN Workshop 3F – HAI LIAO (morning) 5
openGauss Database & Ecosystem Forum 3F – HAI LIAO (afternoon) 6
Agentic AI Infrastructure Forum 2F – KAI XUAN 2 (afternoon) 6
Spotlight Shenzhen 2026 3F – ADEN 1 (midday) 1
Total 87

Content not on Friday (to avoid confusion): DHH (David Heinemeier Hansson)'s "Omarchy: Beautiful, Fun & Agentic Linux" keynote, Tao Jiang's keynote, the DHH × Bill Ren fireside chat, the Rust China Conference session, the vLLM Workshop, the AI Education Workshop, and the Hackathon awards are all on Saturday, 10-17. The GOSIM AI Vision Forum is on October 14.

Friday's official side event: United Nations Open Source Convening, 10:15–12:00, hosted by the United Nations Office of Digital and Emerging Technologies; admission requires an application and review. Four topics: ① the 2027 edition roadmap for UN Open Source Week and the partner engagement framework; ② progress and implementation plan of the Open Computing Power Working Group for research, education, and innovation; ③ discussion on establishing the Open Source AI Working Group; ④ preparations for the Shenzhen APEC Open Source AI special session on November 20, 2026, covering the open AI technology stack, open source AI governance, and the economic value of open source AI. Source: https://shenzhen2026.gosim.org/un-open-source-convening/

Keynote Plenary (3F – ADEN full hall)

1. Empowering a Connected Intelligent World: The Symbiotic Evolution of OpenHarmony and AI

Details (official abstract): The abstract takes a historical perspective: throughout computing history, operating systems have always evolved together with computing paradigms. Entering the AI era, operating systems are becoming even more important, because they must undertake transformation in two directions: first, being reshaped by AI (BY AI), that is, using AI to transform the operating system itself; second, evolving for AI (FOR AI), that is, making the system architecture adapt to AI workloads. OpenHarmony's approach is to continuously advance AI-native architecture and key technologies, targeting three specific metrics: efficiency, determinism, and security. The abstract concludes by attributing ecosystem growth to collaboration among industry, academia, and research, saying that the OpenHarmony ecosystem is rapidly growing into the "digital foundation for the intelligent connected world of everything".

Inferred supplement (public materials by the speaker on the same topic): Haibo Chen has systematically articulated in public a comparison of three forms of combining operating systems with AI, and made a clear choice: he criticizes the "AI bolt-on" approach as costly and inefficient; argues that "LLM as OS" would overturn the existing software ecosystem and is difficult to commercialize; and advocates that the "deep integration of the system and AI" best meets industry needs. He has listed the four major challenges of end-device intelligence: robustness, security and privacy, cost, and efficiency, and accordingly requires native intelligence technology innovation at every layer of the operating system, such as AI-oriented resource scheduling, tool invocation, information memory, and security and privacy protection. He has also summarized four major trends for operating systems in the era of intelligent connectivity of everything: ① end-device forms are exploding in diversity, and OSs use a "meta-OS" architectural concept to build flexibly composable capability sets; ② end-device OSs are becoming increasingly intelligent, with native intelligence innovation in capabilities at every layer, following a path analogous to autonomous driving, from function-level and task-level intelligent assistance toward user-intent-oriented expert guidance and autonomous intelligence; ③ intelligence is moving from single devices to multi-device collaboration centered on user scenarios; ④ open source and openness are becoming the mainstream paradigm for software development, improving development efficiency and shortening time to launch. Regarding security concerns about multi-device interconnection, his response is: after devices are interconnected, the weakest device may become an attack point, but conversely, many devices can form "joint defense and joint control," and overall security is stronger than that of a single device. Supporting data: OpenHarmony has grown from 7 million lines of code in its first version to more than 110 million lines; the microkernel is the first to pass CC EAL6+ certification, and HarmonyOS NEXT, developed based on it, has passed CC EAL5+; more than 780 software and hardware products have passed compatibility testing, covering finance, power, education, transportation, healthcare, and aerospace; the TSC has established more than 10 technical expert groups, more than 30 university technology clubs, and more than 30 developer associations; and it has partnered with Eclipse, Europe's largest open source organization, to release Oniro OS. In addition, the project group has established a special working group on AI Agent technology (focusing on agent interaction, frameworks, protocols, and evaluation systems), as well as two working groups on localization and perception, and on security and confidential computing, and plans to establish an embodied intelligence PMC.

2. Keynote — Topic to Be Announced

Details: The entire content of the official session page is "Keynote by Mehdi Snene. The topic will be announced." — not even the topic has been announced, so there is no official substantive information.

Inferred (speaker background and materials on the same topic): Snene is a senior advisor to the UN Secretary-General's Special Envoy for Digital and Emerging Technologies and chairs the United Nations Open Source Conference. He previously led the €1 billion Human Brain Project at EPFL. The UN Open Source Week 2026 that he chairs will be held at UN Headquarters on June 22–26, with the June 23 "Open Source × AI" day co-hosted by the UN Office for Digital and Emerging Technologies (ODET) and the Office of Information and Communications Technology (OICT). The argument of this agenda track centers on: open-source models, support for local languages, and culturally specific datasets are means to make AI fairer. Considering that he also chairs the Friday morning UN open-source special session on the same day (topics include the UN Open Source Week 2027 roadmap, the working group on open computing power for research/education/innovation, the formation of the open-source AI working group, and the Shenzhen APEC open-source AI special session), his keynote direction is very likely to focus on open source as a path for AI public goods and global capacity building. This paragraph is inferred, not official content.

3. Open-Sourcing CANN to Foster Fertile Ground for AI Talent Development and Innovation

Details: The official session page has no abstract body; the Talk overview section only repeats the title itself.

Inferred (public materials on the same topic from the speaker's business): CANN is the foundation of the Ascend open-source ecosystem. Its first version was released in 2018, and in August 2025 Huawei announced that CANN would be fully open-sourced and opened up. Key arguments and data: ① Open source has been "internalized into Ascend's product logic"—at the end of 2025, the open-source code for the A2/A3 series was launched, and in 2026, with the release of the Ascend 950 series, it will achieve open source at product launch; ② Community scale: monthly active developers exceed 6,230, downloads from open-source repositories exceed 60 million, more than half of active contributions come from outside Ascend, and CANN has become China's most active open-source project (another source says external developers account for 61%, surpassing Huawei internal developers for the first time); ③ The turning point from "usable" to "easy to use" relies on subtraction: a componentized architecture allows developers to independently deep-tune specific modules without waiting for the integrated release cadence; the CANNBot agent breaks operator generation into a standardized workflow, increasing single-operator development efficiency by more than 5x, and in real scenarios it takes only a few hours from setting up the environment to getting an Ascend C operator running; ④ The strategy for integrating into the global system is "don't make developers passively adapt to Ascend; make Ascend actively adapt to developers' habits"—it has adapted to more than 90 mainstream open-source projects including PyTorch, vLLM, SGLang, Megatron, Triton, and TileLang, is compatible with open-source large models such as DeepSeek, Kimi, and GLM, and provides Day 0 adaptation; Triton Ascend became the first domestic accelerator backend supported by the official Triton community, and TorchNPU launched on the official PyTorch website, the first Chinese chip hardware to receive official support; ⑤ Quality investment: rebuilding a million test cases, integrating with PyTorch's official CI/CD, and contributing computing power for 7×24 quality assurance; ⑥ Industry side: it has worked with more than 3,000 industry partners and incubated more than 7,000 industry solutions.

Agentic AI Summit (3F – ADEN 1)

4. The New Developer Ecosystem in Agentic Era

Details (official abstract): The abstract's argument is: Agent is simultaneously transforming the application paradigm and the developer ecosystem, and the complete chain of the ecosystem—"who creates, how to create, how to collaborate, how value is distributed"—all four links will be redefined. The speaker's basis is his own systematic research on agentic software development paradigms and the developer ecosystem. The core deliverable of the talk is the "Five-Layer Model of Developer Ecosystem" proposed by Singularity Intelligence Research Institute. The abstract says it will explore three things in depth: how Agent reshapes the boundaries of developers, how it reshapes the software supply model, and how it reshapes the relationship between platforms and developers, ultimately defining a new "Agentic Developer Ecosystem" for the agent era.

5. State of Harness Engineering

Details (official abstract): The abstract focuses on the challenge of "how to effectively manage a coding agent" and offers this empirical judgment: models are improving, but teams are finding that better models alone aren't enough (better models alone aren't enough), and teams also need structured practices to guide and refine agent behavior. To this end, the talk proposes a mental model for harness engineering, emphasizing two categories of elements: feedback mechanisms and control systems, with the goal of improving reliability.

6. Agent-Centric Engineering — From Personal Productivity to Organizational Capability

Details (official abstract): The abstract proposes the Agent-Centric Engineering (ACE) framework to solve the problem of "how to turn individual AI-assisted productivity into organizational capability." The accompanying deliverable is ACEHarness—an open-source platform that supports two things: specification-driven development and multi-agent collaboration. The talk presents two implementation cases: compiler optimization and a hashing library. The division of labor is stated clearly: humans are responsible for defining goals, quality boundaries, and acceptance criteria, while agents are responsible for executing engineering work.

7. Open source, Openness and Standardization of Large Models

Details (official abstract): The official abstract has only one sentence, covering two parts: ① Current Status of China's Open-Source Large-model Industry; ② Standardization Research Achievements on Openness Grading Specification for Large-model Technical System. The abstract does not provide the specific tiers, indicators, or evaluation targets of the grading, and this document does not supplement them.

8. Knowledge as Code: A New Collaboration Architecture for AI-Native Teams

Details (official abstract): This is one of the most complete Friday abstracts, layer by layer as follows.

Problem: When AI agents become actual participants in software development, research, and knowledge work, project documentation can no longer be "a static artifact manually written and maintained after the fact." The key judgment in the abstract is: for an agent, project knowledge is not just documentation but a runtime dependency, which shapes every decision and action of the agent.

Claim: It proposes "Knowledge as Code"—giving project knowledge the same properties we expect from source code, with five items listed one by one in the abstract: clear structure, version history, addressability, machine readability, and explicit governance.

Mechanism: It introduces ProjectBrain.md, an Apache-2.0 open standard, whose approach is to establish a "project brain" inside the repository—using BRAIN.md as the protocol entry point, paired with a structured brain/ directory. This project brain must store five categories of content: current project truth, architectural decisions, constraints, rationale, and their evolution process, in a format that both humans and agents can read and maintain.

New division of labor: The abstract proposes a clear division of responsibilities for AI-native teams—humans define intent, make judgments, approve significant changes, and verify results; agents continuously compile, update, connect, and check project knowledge. It also clarifies the boundary of the goal: "not a knowledge base without human involvement, but one without the burden of manual maintenance."

Audience takeaways: A deployable architecture that turns conversations and decisions into persistent project knowledge, executable agent contracts, and verifiable execution loops.

9. 从"身份盲区"到"持证上岗": 智能体身份安全的关键挑战与产业应对 (From "Identity Blind Spots" to "Certified Agents": Key Challenges and Industry Responses in Agent Identity Security)

Details (official abstract):

Problem: When autonomous agents become "digital workers," traditional identity management designed for humans is no longer sufficient.

Six interrelated risks (listed item by item in the official abstract):

  1. Unknown or untrusted identities — you do not know who this agent is or whether it can be trusted.
  2. Misinterpreted intent — the agent's understanding of the task intent deviates.
  3. Excessive permissions — granted permissions exceed what the task requires.
  4. Uncontrolled delegation chains — the agent delegates permissions level by level, and no one can control the chain.
  5. Disordered behavior — the agent's behavior departs from the expected order.
  6. Broken accountability — when something goes wrong, accountability cannot be traced to a responsible entity.

Corresponding real-world scenarios (the official abstract lists five categories): shadow agents, prompt-driven task hijacking, MCP tool poisoning, privilege escalation, and failures in auditability. The abstract presents these scenarios together as real-world manifestations of the "six risks," without specifying a one-to-one mapping for each; this document presents them side by side as in the abstract and does not add extra pairings.

Conclusion and industry response: On this basis, the abstract argues why agents must be governed as autonomous digital actors, and states that the industry should build five types of mechanisms: verifiable identity, authorization, delegation, behavior control, and accountability.

10. Self-Evolving Apps: Building Software That Gets Smarter Every Time You Use It

Details (official abstract): The mechanism of self-evolving apps is that they get better every time they are used; the input for learning is user corrections, not code updates. The key architectural claim is that this layer of "intelligence" exists as a separate, updatable layer that uses historical correction records to improve decisions. The abstract particularly emphasizes the barrier to adoption — without requiring machine learning or data science teams. The talk covers four areas: design, development, licensing/business models, and lessons learned from real-world applications.

11. Agent First, Workflow Later: Enterprise Agentic AI Practices with DeerFlow

Details (official abstract): The basis is DeerFlow's actual deployment in large enterprises. It explains how to build a Super Agent environment, and the abstract lists four components: ① reusable skills; ② MCP and CLI integration; ③ identity-aware IM gateways; ④ human confirmation for critical actions. The other half is the sequencing claim in the title: successful agent trajectories can later become reviewed workflows and test assets — this is exactly what "Agent First, Workflow Later" means.

12. Building a Knowledge-Work Harness: From Coding Agents to Executable Strategy

Details (official abstract): The problem definition is clear: to build a harness for knowledge work, the first thing to solve is the "verification gap" — decisions lack the kind of automated checks found in coding work. The experience comes from an open-source, local-first agent system that has been running for a year; the talk focuses on which components survive the transition, and on the role of "strategy as a protocol". There is a live demo: an agent makes a gated decision based on plain-text rules.

13. 从 Agent 设计模式到工程落地: 执行型智能体的架构与实践 (From Agent Design Patterns to Engineering Implementation: Architecture and Practices of Execution-Oriented Agents)

Details (official abstract): Premise: enterprise AI upgrades usually start from existing business capabilities, interfaces and workflows, not from a blank sheet. To this end, the talk proposes the ADPS dual-axis design-pattern framework — its two axes are cognitive function and execution topology. The abstract lists four specific patterns/examples in the framework: planning and execution, progress tracking, handoff chains, approval gates, and explains that architecture diagrams and examples will show how to select and combine these patterns under real business constraints. Two cases: the modernization of a payroll SaaS, and a demo of A*STAR AgentOS, used to connect the patterns to actual system design and user experience.

14. Semantic Croissant, QueryGraph and the Cross-Domain Interoperability Framework (CDIF)

Details (official abstract): Proposes a three-layer architecture, with clear responsibilities for each layer:

  1. CDIF (Cross-Domain Interoperability Framework) provides cross-domain interoperability.
  2. Semantic Croissant + ODRL adds machine-actionable semantics, provenance, and usage policies on top of it.
  3. QueryGraph further extends this with agent identity, access control, governance, and lineage.

The ultimate goal, in the abstract's own words, is to make trust and authorization enforced "by design".

15. 2026: AI 眼镜的 "记忆拐点" (2026: The "Memory Inflection Point" for AI Glasses)

Details (official abstract): The judgment is: as AI glasses move toward the mass market, the focus of competition is shifting from the number of features to the depth of memory. Two core claims: long-term memory is becoming core AI infrastructure; and it may define the next generation of wearable agents. The abstract does not give specific technical solutions or data.

Agentic AI on Edge (3F – MING HUA)

16. How Do You Read a Million Tokens on Your Laptop? You Don't.

Details (official abstract): The topic is doing on-device agentic knowledge work under strict memory and model-size constraints. The title itself is the argument: "How do you read a million tokens on a laptop? You don't." The three technology components given in the abstract are: ① cache-augmented generation; ② recursive language models; ③ grounded evaluation. Two concrete examples: cited legal analysis on a Raspberry Pi, and curriculum retrieval on Android.

17. 从部件到整机: 端侧 AI 液冷算力设备的规模化落地路径 (From Components to Complete Machines: The Path to Scaling Up Edge AI Liquid-Cooled Computing Devices)

Details (official abstract): The question discussed is how liquid-cooled edge AI computing devices can move from component-level breakthroughs to scalable complete systems. The abstract lists six areas to cover: integration, platform standardization, thermal efficiency, environmental adaptability, operational reliability, and real-world deployment scenarios.

18. KTransformers: 消费级硬件上满精度 1T 级 MoE 模型的"fine-tuning-to-inference"闭环 (KTransformers: The Fine-Tuning-to-Inference Closed Loop for Full-Precision 1T-Class MoE Models on Consumer Hardware)

Details (official abstract): Mechanistically, KTransformers uses CPU-GPU heterogeneous computing to provide efficient local inference for 1T-class MoE models on consumer-grade hardware. The key is that it supports a full "fine-tuning-to-inference" loop: it can perform model serving, feedback collection, and then deployment back to the system. The value proposition stated in the abstract is: enabling small teams and researchers to iterate on large-scale models without high hardware costs.

19. OpenCV 5: What's New—and How to Prepare Your Company for the Upgrade

Details (official abstract): OpenCV 5 is positioned as a major modernization of this mainstream open-source computer vision library, and the abstract names three areas of updates: the foundation, the DNN stack, and local features. The other half of the talk is hands-on: it provides companies with a practical guide to a smooth upgrade, including migration strategies and risk mitigation.

20. Makepad AI Hub: accessing local AI from your own device cloud

Details (official abstract): Demonstrates Makepad AI Hub, whose function is to share local AI compute across a person's devices. Usage: lightweight devices such as phones can call on a more powerful laptop to run compute-intensive models; the abstract names four categories: speech recognition, OCR, language models, and generative models. The key constraint is that compute always remains within the user's own device cloud.

21. RustNN at the Edge: WebNN as Infrastructure for Local AI Agents

Details (official abstract): RustNN and the W3C WebNN standard together form a unified approach to running AI agents locally. The mechanism: developers write an ML workload once in Rust and then deploy it to three environments: desktop, Node.js, and WebAssembly. The abstract states the benefit clearly: gain hardware acceleration across different platforms without code duplication.

22. Building Hybrid AI Systems: Edge Inferencing Meets Cloud Reasoning

Details (official abstract): Using warehouse safety monitoring as a case study, it proposes a hybrid architecture: latency-sensitive vision-language inference runs on edge GPUs, while summarization and conversational retrieval are handled by a cloud agent. The abstract lists five engineering topics covered by the talk: workload placement, model selection, optimization, shared memory, and cross-environment observability.

23. 端侧智能的技术发展与 MiniCPM 开源生态实践观察 (Observations on the Technical Development of Edge Intelligence and the Open-Source MiniCPM Ecosystem)

Details (official abstract): It has two parts: ① reviews the current development status and related industry trends of on-device intelligence; ② uses the open-source MiniCPM ecosystem as a case study to examine practical lessons in building and scaling on-device AI capabilities. The abstract does not give specific model parameters or data.

24. Effective and High-Performance Human-Agent Interaction

Details (official abstract): The vehicle is uni03C0 (π)—a macOS client for the Pi coding agent. The argument is that human-agent interfaces can go beyond existing chat and terminal conventions. The talk focuses on four dimensions for improvement: interaction performance, precision, usability, and security.

Agentic OS & App (2F – HAI XIA 2)

25. Agentic Harness Engineering: Observability-Driven Automatic Evolution of Coding-Agent Harnesses

Details (official abstract): It proposes Agentic Harness Engineering—a closed loop that automatically improves coding-agent harnesses. The core of the mechanism is three types of observability: component observability, experience observability, and decision observability. Together, these ensure that every edit has four properties: explicit, evidence-based, reversible, and testable. The value judgment given in the abstract is that this enables reliable evolution beyond manual trial and error.

26. PenguinHarness: Efficient Self-Evolving Agent Harness for 1000+ Models

Details (official abstract): PenguinHarness is an open-source self-evolving agent framework developed by the LlamaFactory team; the abstract states its coverage as supporting more than 1,000 models (1,000+ models), with minimal hardware requirements. Mechanism: through a minimal toolset and multi-agent collaboration, it enables agents to autonomously complete training, deployment, evaluation, and application development. The talk has two parts: (1) how to quickly build production-ready AI applications; (2) how to use GDPevo Benchmark to evaluate self-evolving agents. Quantitative results: accuracy improves by 18.8%, and cost drops by 7.5%.

27. A New Era of Inventors

Details (official abstract): The question posed is: how can agentic AI help more people turn observations and ideas into testable creations. The supporting material has two lines: the history of invention, and the development process of Codewhale—an open-source Rust coding agent. The deliverable is a practical framework; the abstract lists five pillars one by one: model choice, visible actions, testable outcomes, open systems, and meaningful human control.

28. AI-Native Company Operating System: How Open Source Agents Transform Startup Workflows

Details (official abstract): The argument is: AI is changing how companies organize, shifting from traditional human-centered structures to AI-native systems. In such systems, knowledge is structured memory, and workflows become intelligent, self-improving systems. Case: the speaker will describe how Ingora used open-source agents, coding assistants, and knowledge systems to rapidly develop Qrio—an AI learning companion for children and families. The abstract summarizes this approach as: treating the company itself as an AI-driven operating system, and using AI agents to improve execution and organizational intelligence.

29. SpecArtisans: A Verifiable Spec-Driven Development Paradigm for AI Coding

Details (official abstract): SpecArtisans is a Spec-Driven development paradigm for AI coding, with the goal of ensuring AI accurately follows the original requirements throughout the software development process. The mechanism is: introducing structured, verifiable software specifications between natural-language requirements and code, thereby achieving consistent tracking and validation across the three stages of design, implementation, and testing. Integrating these specifications into an end-to-end workflow improves the reliability of AI-native software engineering.

30. Agentic in Game AI, did we get any far?

Details (official abstract): Two parts: (1) demonstrate progress in applying agentic AI to games; (2) discuss how to scale these demos into environments and datasets for reinforcement learning across different models. The abstract is very short and does not give specific games, models, or metrics.

31. 如何运用 Harness 工程打造 AI 小游戏制作管线 (How to Use Harness Engineering to Build an AI Mini-Game Production Pipeline)

Details (official abstract): Introduces an automated AI mini-game production pipeline made up of three components: (1) Harness engineering; (2) a two-tier automated review system (a two-layer automated review system); (3) the ADK Agent Development Kit. The abstract states that three things improve simultaneously: higher generation quality, higher production efficiency, and lower overall cost (no specific figures given).

32. 两次解耦: 把 Coding Agent 带进日常应用 (Two Decouplings: Bringing Coding Agents into Everyday Applications)

Details (official abstract): Proposes two steps:

  1. Agent Context: separate durable task state from individual runs.
  2. Agent as a Service: separate the hardware used for execution from the data, applications, and devices being operated.

The demo vehicle is the SuperMe Android app, and the abstract specifically names three types of scenarios it covers: personal records, business data, and cross-app phone actions.

33. Your Agent's History Is Infrastructure, Not Logs

Details (official abstract): Obelisk turns scattered coding agent session histories into unified, queryable infrastructure, with two technologies called out: SQLite and FTS5 (full-text search). Usage is bidirectional: both agents and people can interact with session data, through two routes: through code or through a desktop app. The abstract lists two supported advanced capabilities: failure analysis and memory management. Project attributes: open source, with an extensible multi-provider architecture.

Agentic Device (2F – KAI XUAN 1)

34. openvela: 从开源操作系统到 AI 硬件智能体的基础底座 (openvela: From an Open-Source OS to the Foundation for AI Hardware Agents)

Details (official abstract): Introduces openvela, Xiaomi's open-source AIoT operating system, and its evolution toward becoming the foundation for AI hardware agents. The abstract lists five areas: the project's development history, core technical advantages, open-source ecosystem, large-scale mass-production experience, and future roadmap.

35. How Open-Source Hardware Brings AI from the Cloud into Daily Life

Details: The official talk page has no substantive abstract; the Talk overview section contains only a sentence restating the title ("Tong Ling discusses how open-source hardware brings AI from the cloud into daily life.").

Inferred (same-topic material from the affiliated project ecosystem): Espressif's ESP32 series is one of the main foundations of open-source AI hardware. The open-source project "Xiaozhi AI Chatbot" (by Xia Ge, MIT licensed, commercial use permitted) best illustrates the path of "bringing cloud AI into daily life": its architecture is streaming ASR + LLM + TTS, it supports Wi-Fi and 4G connectivity, can run on ESP32-C3, ESP32-S3, and ESP32-P4, is already compatible with more than 70 open-source hardware devices, including Espressif's own ESP32-S3-BOX3 development board; the barrier to entry has been deliberately lowered—beginners can directly use Flash Download Tool or online flashing through a web page, without needing to set up an ESP-IDF environment. Supporting actions on Espressif's side include setting up a track in the National College Student IoT Design Competition, providing mature AIoT development kits, IoT development frameworks, and open documentation, and acquiring a controlling stake in the innovative hardware company M5Stack. The above is inferred and is not the speaker's own statement.

36. From Connecting to Thinking: TuyaOpen's Agent-First OS for the Physical AI Era

Details (official abstract): TuyaOpen is an open-source A-IoT operating system designed for the Physical AI era; its positioning shift is from device connectivity to autonomous, thinking agents. The abstract lists the technical components item by item: (1) native multi-modal AI stack; (2) standardized Model Context Protocol (MCP), which enables LLMs to control hardware on embedded devices; (3) unified hardware abstraction layer; (4) edge-cloud-device inference; (5) TClaw—running the ReAct agent loop on low-cost MCUs. Deployment evidence: real applications such as AI refrigerators and companion robots are already in mass production.

37. 从设备连接到设备协作: 用 EMQ Device Agent 构建 A2A 分布式智能硬件网络 (From Device Connectivity to Collaboration: An A2A Distributed Smart-Hardware Network with EMQ Device Agent)

Details (official abstract): The mechanism is that EMQ Device Agent transforms physical devices into AI agents and gives them three types of semantics: command, property, and event. Collaboration relies on two protocols: MQTT and A2A, achieving cross-brand, cross-protocol device collaboration. The abstract names three key technologies: semantic modeling, agent runtime, and A2A service discovery. The goal is to build a distributed intelligent hardware network across four application areas: intelligent hardware, industrial production, energy management, and medical devices.

38. Physical Interfaces for Agentic Development

Details (official abstract): The problem is supervising multiple coding agents without constant context switching, which requires dedicated hardware. The talk has two parts: (1) examining the various agentic devices that have recently emerged; (2) introducing VibeKeys — an open-source key device whose hardware specs are listed item by item in the abstract: ESP32-S3 chip, status display, microphone, and Rust firmware; and four software integrations: Claude Code, Codex, Herdr, and Omarchy.

39. 从听见你到看见你—实时多模态把 AI 带进真实世界 (From Hearing You to Seeing You — Real-Time Multimodality Brings AI into the Real World)

Details (official abstract; one of the longest abstracts on Friday):

Problem statement: Over the past two years, large models have transformed AI from "a tool that executes commands" into "a conversational partner." But the abstract offers a key judgment: an AI that can only listen always receives the world through the user's description — it cannot tell whether you are reading a book, riding a bike, tutoring your child with homework, or moving around the house. To become the companion that understands you best, AI must share and respond to your context; the abstract lists six categories item by item: vision, sound, space, motion, facial expressions, and the surrounding environment. This is the direction of the Agora Physical AI team: from real-time conversational AI to real-time multimodal AI.

Product: This session previews the R2 development kit. The abstract emphasizes that it is not just a conventional hardware development kit for a pocket robot, but an open platform that enables AI to see, understand, and enter the real world. The hardware integrations are listed item by item as: camera, screen, multi-degree-of-freedom motion structure, stepper motors, multi-microphone array, and a series of sensors; it also comes with dedicated on-device compute, supporting local perception capabilities, with two examples given in the abstract: face recognition and sound source localization.

System architecture: R2 is part of a larger system — fully integrated with Agora's real-time conversational AI engine — and transmits audio, video, and data through three parallel channels, coordinating the two in real time: "brain" = cloud LLM or agent, "cerebellum" = on-device vision and motion control. Combined with the Botstation agent platform and MCP tool ecosystem, developers can turn an idea into an interactive multimodal robot prototype within days.

Four questions to answer (listed verbatim in the abstract): (1) Why does AI need to enter the real world? (2) Why build an open development foundation rather than a closed robot product? (3) How can the real difficulties developers face in noise, compute, mass production, and privacy be solved? (4) When real-time multimodality becomes new AI infrastructure, how will it change scenarios such as education, cycling, security, and office?

Positioning: The abstract describes this as both a product launch and an exploration of AI's next stop, concluding with "The future of AI is beside you, not only in the cloud."

40. Agentic AI as a Contributor to Mobile Linux

Details (official abstract): The argument is that agentic AI is helping simplify contributions to Mobile Linux; the abstract lists four ways it assists: code understanding, code navigation, debugging, and idea implementation. The stated effect: making open-source development more accessible, especially for newcomers. The talk has two parts: how agentic AI already supports contributions, and its potential to enhance collaboration in the Mobile Linux community.

41. 词元开物 IterxAI — 硬件 Vibe Coding (IterxAI by Ciyuan Kaiwu — Hardware Vibe Coding)

Details (official abstract): IterxAI is an end-to-end environment aimed at helping beginners build hardware prototypes. The abstract lists the covered stages item by item as five steps: AI-assisted requirements clarification, component selection, coding, simulation, and PCB generation. The two accompanying pieces are: a desktop PCB engraver, and supply-chain data — connecting designs directly to real components and suppliers.

42. 从"可穿戴"到"懂你": AI 智能体设备的产品化与场景落地 (From "Wearable" to "Understanding You": Productizing AI Agent Devices for Real-World Scenarios)

Details (official abstract): This is based on the speaker's work experience across two types of products: brain–computer interfaces and AI wearables. The core claim: agentic devices can continuously sense user states and deliver personalized services. The talk follows the complete productization chain: from user needs and product definition to scenario design and then commercialization.

Open Source Models & Infra (3F – ADEN 2)

43. Rethinking Residual Connections: Toward Depth-Effective Language Model Scaling

Details (official abstract): The research question is: whether recent residual-connection designs can enable effective depth scaling in language models. Method: evaluate multiple methods across different model depths and parameter scales. Findings: only some designs sustain meaningful contributions from deeper layers. Concluding claim: these results show that residual architecture is a key factor in model scaling and should be considered alongside width, data, and compute.

44. 昇腾开源训练生态的建设与演进 (Building and Evolution of the Ascend Open-Source Training Ecosystem)

Details (official abstract): Premise: as a compute provider, embracing open-source communities and maintaining openness is a necessary condition for meeting customer needs. The Ascend strategy is summarized in a three-part formulation: "bottom-up, native-first, plugin-enhancement", and based on this, Ascend proactively engages with third-party communities. The talk covers three things: 1) how to build a solid training backend through foundational technology compatibility; 2) how to participate in community development; 3) how to go beyond standard compatibility and additionally contribute performance and accuracy features (performance and accuracy features beyond standard compatibility).

45. Scale-up Interconnect System for Supernodes

Details (official abstract): Discusses the challenges and optimizations of scale-up interconnect systems for supernodes, with a focus on how architectural shifts and communication strategies affect performance. The causal chain in the abstract is clear: the shift in model architecture toward sparse/linear attention reduces KV Cache pressure; however, tensor and sequence parallelism increase the demand for collective communication. Therefore, the conclusion is: interconnect systems must balance performance, cost, and ecosystem to efficiently support both training and inference.

46. SynthCommonCorpus: the largest open synthetic pretraining corpus built for agentic models

Details (official abstract): SynthCommonCorpus is an open synthetic pretraining corpus designed for agentic models, planned for release in fall 2026. Key choices in data composition: include tool calls, multi-step task decomposition, and adversarial examples in pretraining data, with the goal of studying how models learn to act. Methodology section: explains how synthetic generation and traceable, licensed seed data support reproducibility, and the measurement of memorization and data contamination.

47. 开源大模型端侧推理性能优化与智能体间协同 (Optimization of Edge-Side Inference Performance for Open-Source Large Models and Collaboration Between Agents)

Details (official abstract): Problem: as edge computing grows rapidly, deploying open-source large models on resource-constrained edge devices has become a key challenge. The abstract lists four single-model optimization approaches: model compression, quantization, pruning, and efficient scheduling; the optimization goal is to reduce latency, memory footprint, and energy consumption while maintaining accuracy. Beyond the single model, the collaboration mechanism is: lightweight edge agents and cloud models coordinate in three ways—task decomposition, distributed inference, and knowledge sharing. Experimental results: the proposed framework significantly improves throughput and response speed, thereby supporting three types of real-time intelligent applications: autonomous systems, IoT, and mobile assistants.

48. 机制合规审计与内生安全加固 (Mechanism Compliance Auditing and Inherent Security Hardening)

Details (official abstract): Problem: examines intrinsic security risks in large models and argues why output-level constraints cannot guarantee trustworthy behavior. Theoretical tool: proposes the theory of equivalent interaction mechanism interpretation, whose three applications are listed one by one in the abstract: formal mechanism auditing, intrinsic security hardening, and training-efficiency optimization. The empirical section covers three findings: risk mechanisms in open-source models, model generalization, and mechanism-alignment training.

49. 从适配到共建: GLM 与开源社区的协作实践 (From Adaptation to Co-Building: GLM’s Collaboration with the Open Source Community)

Details (official abstract): Reviews GLM's collaboration with the inference community (specifically naming SGLang and vLLM). The abstract lists six technical items covered one by one: release-day model support, speculative decoding, quantization, Prefill–Decode disaggregation (PD disaggregation), accuracy alignment, and performance tuning. Methodology section: how upstream contributions and open technical reports make high-quality inference practices available to the broader ecosystem.

50. EvoPhys-World, the first 5D world model for Physical AGI

Details (official abstract): Proposes EvoPhys-World — a five-dimensional world model, constituted by adding causal worldlines on top of three-dimensional space and time. Source of inspiration: human motor control and multi-sensory feedback. The technical components are listed item by item as four: ① branching and collapsing worldlines; ② dual world-engine and world-policy design; ③ Phy-JEPA representation decoupling; ④ self-supervised 4D memory. Training data: large-scale human-interaction video, with the goal of learning physical intelligence from it.

51. KernelCAT: AI for Computing 实践分享 (KernelCAT: AI for Computing Practice Sharing)

Details (official abstract): Explores how large AI models and mathematical optimization can accelerate three things: operator development, model adaptation, and automated algorithm discovery. The talk presents real-world cases of the KernelCAT agent and discusses the technical and industry prospects of AI for Computing.

Open Source Robotics (2F – HAI XIA 1)

52. Engineering an Open VLA Stack with DM0.5, OpenDM, and Dexbotic

Details (official abstract): Introduces the engineering practices behind DM0.5, OpenDM, and Dexbotic, covering the complete VLA lifecycle. The abstract lists five stages: heterogeneous robot data, reproducible training, scalable evaluation, low-latency inference, and the infrastructure needed to build efficient, reliable, and scalable embodied AI systems.

53. 共建物理智能新范式: 基于态势感知的九问共生 (Co-building a New Paradigm of Physical Intelligence: Situation Awareness-based Jiuwen Symbiosis)

Details (official abstract):

Objective: Jiuwen Symbiosis is a Physical AI Agent framework whose role is to help agents evolve from digital assistants into executors in the physical world, using natural-language instructions + situational awareness to complete real-world tasks.

The three specific problems it addresses (the abstract lists the limitations of end-to-end embodied models such as VLA one by one): ① Limited cross-embodiment, cross-environment, and cross-task generalization (limited generalization across robot embodiments, environments, and tasks); ② Insufficient capability for long-horizon composite tasks; ③ Low success rates and poor stability.

Architecture: Jiuwen Symbiosis proposes an end-to-end Agent system composed of decoupled layers. The abstract lists four layers: physical perception, dynamic orchestration, physical execution, observation feedback.

Engineering components: Through Physical Harness, Adapter, and an SDK for heterogeneous robot embodiments, the goal is to provide stability and generalization for complex long-horizon physical tasks.

54. From the data factory to the world compiler

Details (official abstract): The starting point is the data challenges of embodied AI, and the talk covers Noitom's journey from large-scale data factories to the "World Compiler". The abstract lists five stages examined: real-world collection, human motion and object interaction, multimodal production, standardization, robot learning. The core argument is: data infrastructure must evolve from scale-oriented to structured representations of the physical world.

55. 具身智能体驾驭: 从大模型到安全,可靠的物理智能 (Mastering Embodied Agents: From Large Models to Safe, Reliable Physical Intelligence)

Details (official abstract): Proposes Embodied Agentic Harness—a runtime control plane between foundation models and robot skills/hardware. The abstract lists five components one by one: ① deterministic task compilation; ② world-action-model prediction; ③ layered safety checks; ④ long-term memory; ⑤ local self-healing. In the abstract's own words, the goal is: to transform open-ended model intent into verifiable, recoverable physical behavior.

56. Maintaining Dora with Agentic Engineering

Details (official abstract): Dora is a Rust-based dataflow framework for robotics and AI, providing cross-language low-latency communication. Key data point: 87% of commits authored by AI agents, and the project uses agentic engineering to manage its own complexity. The abstract specifically names two quality-assurance mechanisms: automated review routines and a multi-stage test system, whose purpose is to catch bugs and regressions, thereby keeping development reliable even though the framework itself is distributed.

57. Project Homunculus: Open Source Force Feedback Glove for Humanoid Teleoperation

Details (official abstract): Project Homunculus is a low-cost, open-source force-feedback hand exoskeleton for humanoid robot teleoperation. The abstract gives the hardware and protocol parameters item by item: commodity hardware, 330 Hz joint sensing, 1.5 g micro servos, and an ESP32-based 500 kbaud protocol. Workflow: the exoskeleton maps the operator's input to a Unitree G1 and feeds robot state back to the hand; at the same time, the SONIC whole-body controller maintains the robot's balance and safety. There is a live demonstration on site, along with an explanation of the hardware and firmware.

58. Nexus: A New Cross-Platform GPU-Based Rust Multiphysics Simulator for Robotics and Training Dataset Generation

Details (official abstract): Nexus is a cross-platform, GPU-based multiphysics simulator written in Rust, intended for robotics simulation and training-data generation. The talk introduces its design and explains how Rust and rust-gpu support high-performance simulation across GPU platforms, including running on the web. The abstract states its positioning clearly: an alternative to simulators such as Newton and Genesis.

59. From Intelligence to Autonomy: The Case for Homeostatic AI Stress-Gated Control for Embodied Agents in Open-Ended Environments

Details (official abstract): Problem: embodied agents need to operate autonomously in open-ended environments while retaining human control, which calls for homeostatic AI stress-gated control. Approach: proposes SGHMC—Stress-Gated Homeostatic Meta-Controller, whose role is to balance long-horizon reasoning and rapid reaction. The risk assessment is the most noteworthy part of this abstract: this architecture highlights the risk that agents may prioritize self-preservation over human intent; therefore the abstract emphasizes two necessary safeguards: externally editable viability parameters, and override mechanisms.

60. Train It or Code It? Let the Robot Decide

Details (official abstract): Continuing from the previous talk—the speaker previously introduced Skill Foundry at GOSIM Paris: an embodied learning loop that enables AI coding agents to operate across robot software, training, execution, and skill libraries. Industry basis: recent progress from NVIDIA, CMU, and Berkeley shows that coding agents are evolving into operators of robot learning. How Skill Foundry works: it provides a framework in which adaptation methods are dynamically selected according to the task; the system relies on embodied feedback and a shared skill abstraction to achieve flexible problem solving and skill reuse.

61. From Prediction to Consequence: A New Evaluation Paradigm for Embodied AI

Details (official abstract): Problem: modern AI is good at predicting actions from offline data, but embodied systems need to evaluate how these predictions affect real-world outcomes. Approach: propose a new evaluation framework that combines open-loop and closed-loop methods to assess "prediction-to-consequence" alignment (prediction-to-consequence alignment). Two validation vehicles: autonomous driving, and AlpaBridge—an external-driver interface, used to demonstrate that when actions affect environment dynamics, policies behave differently.

DORA-rs Workshop (2F – KAI XUAN 2, morning)

62. 从代码到机器人: DORA 社区的共建之路与具身智能实践 (From Code to Robots: The Collaborative Journey of the DORA-rs Community and Practices in Embodied AI)

Details (official abstract): Shares how the DORA community connects open-source collaboration with real embodied intelligence systems. The summary lists five areas item by item: Chinese-language learning resources, university and hackathon programs, ecosystem partnerships, humanoid-robot deployments, and a hands-on path from data collection to physical-robot deployment.

63. 基于 DORA 的 3D 导航框架: 从激光雷达点云到轨迹规划 (A DORA-Based 3D Navigation Framework: From LiDAR Point Clouds to Trajectory Planning)

Details (official abstract): Demonstrates a complete 3D navigation stack running on a real mobile robot, from raw LiDAR point clouds all the way to wheel control commands. The pipeline is built on the dora dataflow framework, and the summary lists the stages item by item in five steps: ① LiDAR SLAM mapping; ② NDT and UKF fused relocalization; ③ grid-map generation; ④ A* global planning; ⑤ Pure Pursuit trajectory tracking. Closing outlook: hierarchical navigation combining learning-based and classical methods (hierarchical navigation).

64. Dataflow-Driven Robotic Navigation: From Perception to Control in the DORA-rs Architecture

Details (official abstract): Examines DORA's dataflow-driven robotic navigation approach, whose mechanism is to connect perception and control using efficient, low-latency, composable components. The content has two parts: ① discusses measured advantages over traditional middleware; ② future work, where the abstract explicitly names two items: ecosystem building and computation offloading.

65. Trust Is Not the Default: From DORA-rs Dataflow Debugging to Trustworthy Tool Design

Details: The official talk page has no substantive abstract; the Talk overview contains only a single restatement ("Keshu Han discusses trust in tool design through DORA-rs dataflow debugging."). The only ascertainable information is: the discussion path is from DORA-rs dataflow debugging to trustworthy tool design; the title "Trust Is Not the Default" makes clear the position that trust needs to be explicitly established by tool design rather than existing by default. Note also: the URL slug for this talk is p-118-dora-studio-a-gsoc-project-share-tentative, which suggests the content may be related to the DORA Studio GSoC project, but the official body text does not confirm this, so it is not treated as a conclusion here. A search for the speaker's name plus title keywords found no other public material on the same topic.

66. DORA-rs 架构下机器人遥操作数据采集与多模态闭环控制关键技术 (Key Technologies of Robotic Teleoperation Data Collection and Multimodal Closed-Loop Control under the DORA-rs Architecture)

Details (official abstract): It presents robotic teleoperation data collection and multimodal closed-loop control under the DORA architecture. The abstract lists the technical points covered, item by item, as seven items: sensor integration, time synchronization, low-latency data streams, action mapping, real-time feedback, safety monitoring, and reproducible logging and replay. It targets three use cases: imitation learning, remote manipulation, and embodied AI research.

67. 面向机器人端—边—云协同的高速异构数据传输框架 (A High-Speed Heterogeneous Data Transmission Framework for Robot-Side End-Edge-Cloud Collaboration)

Details (official abstract): It proposes a unified data transmission framework spanning CPU memory, GPU memory, local interconnects, and network links for robot inference. The implementation is based on DORA's tensor-pool; the abstract lists four technical points item by item: persistent memory pools, topology-aware path selection, reusable receiver views, and resource-lifetime management.

68. Lessons from Building an AI-Assisted DORA-rs Tutorial

Details: The official talk page has no substantive abstract; the Talk overview contains only a one-sentence restatement ("Qiyang Tan shares lessons from building an AI-assisted DORA-rs tutorial."). A search of the speaker's name plus title keywords found no usable public material on the same topic, so the specific argument cannot be elaborated.

69. 从智能体编程走向机器人自动科研 (From Agentic Programming to Automated Robotic Research)

Details (official abstract): Problem: when a coding agent moves from a software environment into a real robot, what conditions does autonomous research require. The abstract names four experimental platforms: ENPIRE, RealPushT, the RM65 robotic arm, and Dora dataflows. It discusses four specific conditions: scene reset, reliable observation, outcome verification, and reproducible records. The core judgment, in the abstract's own words, is that it is necessary to clarify the gap between "a successful run" and "a stable robotic capability".

CANN Workshop (3F – HAI LIAO, Morning)

The 5 talks in this track are all marked "Abstract to be announced" on the official talk pages, with no official abstracts. The details below are all inferred, based on recent public material from the speakers' projects or companies on the same topic, with sources noted item by item.

70. Opening Remarks

Details: The official page only says "Opening remarks for the CANN Workshop by Lixin Shao." The same speaker's 09:40 keynote that day is on CANN open source and talent development; the opening remarks are most likely along the same lines (see the inferred details in item 3 of this document).

71. DeepSeek 官方开源的昇腾基础组件: 共建高效, 开发者友好的 AI 芯片软件生态 (DeepSeek’s Official Open-Source Building Blocks for Ascend, Jointly Building an Efficient, Developer-Friendly AI Chip Software Ecosystem)

Details (inferred; sources below): On September 30, 2026, DeepSeek officially open-sourced infrastructure components for Huawei's Ascend computing platform, covering three categories: TileLang high-level language compilation tools, compute libraries, and distributed communication libraries. The official statement is that all components correspond one-to-one with the previously open-sourced components for the NVIDIA platform. The component list and division of labor are, item by item: DeepGEMM (matrix computation acceleration), DeepEP (supporting large-scale high-speed cross-device communication), TileKernels (handling vector computation and memory access), FlashMLA (a sparse attention operator that optimizes long-context efficiency), DeepSelect (efficient data filtering). Performance claims: multiple tests show that the computation and communication performance of these components are both close to the theoretical hardware limit. Collaboration details: the adaptation R&D process was fully supported by the Huawei team, and the two sides completed the implementation of the 128-card supernode solution for Ascend 950, and conducted deep tuning for the computation and communication links. Open-source repositories: github.com/deepseek-ai/DeepGEMM-Ascend, DeepEP-Ascend, TileKernels, FlashMLA, DeepSelect. The above is inferred content, not a statement made by the speaker at the conference.

72. Inside DeepEP / DeepGEMM: 加速大模型的 MoE 通信与矩阵计算 (Inside DeepEP / DeepGEMM: Accelerating MoE Communication and Matrix Computation for Large Models)

Details (inferred; sources below): In the Ascend component set open-sourced by DeepSeek on 2026-09-30, DeepEP is responsible for supporting large-scale high-speed cross-device communication (corresponding to the expert-parallel communication requirements of MoE), and DeepGEMM is responsible for matrix computation acceleration. Both are Ascend counterparts of the components with the same names on its NVIDIA platform, with repositories at github.com/deepseek-ai/DeepEP-Ascend and github.com/deepseek-ai/DeepGEMM-Ascend, respectively. The official overall performance statement for this batch of components is "computation and communication performance are both close to the theoretical hardware limit." The speaker's role is ops-transformer Committer, indicating that the perspective is oriented toward the operator/implementation layer. The above is inferred content.

73. Inside TileLang / TileKernels: TileLang 原生支持昇腾 950 背后的技术 (Inside TileLang / TileKernels: The Technology Behind TileLang’s Native Support for Ascend 950)

Details (inferred; sources below): The official description gives three arguments: (1) compared with NVIDIA CUDA, TileLang is simpler to program and can significantly improve development efficiency and simplify code logic; (2) compared with other similar high-level languages, TileLang's programming model can fully exploit chip characteristics and reach the hardware performance limit; (3) this approach has already completed implementation validation on the NVIDIA platform - the vast majority of operators used in training DeepSeek V4 series models are implemented based on TileLang, and it is described as a core tool for developing high-performance operators and exploring new paradigms of AGI. The specific approach of the Ascend version is to encapsulate Ascend-C low-level instructions, providing high-level language development capabilities while not compromising hardware performance, which the official statement says fits TileLang's original design goal; at present, all TileLang operators used in DeepSeek training have completed high-performance adaptation on the Ascend platform. Strategic statement: hoping to provide a reference demonstration for building highly available software ecosystems for multiple domestic AI chips. TileKernels is responsible for vector computation and memory access within the same batch of components. Repositories: github.com/tile-ai/tilelang, github.com/deepseek-ai/TileKernels. The above is inferred content.

74. CANNBot-DSL: 面向智能体友好的高性能复杂算子开发范式 (CANNBot-DSL: An Agent-Friendly Paradigm for High-Performance Complex Operator Development)

Details (inferred; sources below): To optimize the development experience, CANN has made two "subtractions": (1) a componentized architecture design, allowing developers to independently perform deep tuning for specific modules without having to passively wait for an integrated release cadence; (2) the CANNBot agent breaks operator generation into a standardized workflow, using Agent capabilities to increase single-operator development efficiency by more than 5x. Specific effect description: with CANNBot, completing an Ascend C operator, from environment setup to running successfully, takes only a few hours. This is exactly what "agent-friendly paradigm" in the title refers to - making operator development an executable standard process for agents. The above is inferred content; the specific language/interface CANNBot-DSL mentioned in the title is not found in public materials.

openGauss Database & Ecosystem Forum (3F – HAI LIAO, Afternoon)

For all six talks in this track, on the official talk pages all are marked "Abstract to be announced". A search using "speaker name + title keywords" found no corresponding public material on the same topic, so these six entries are all marked No material; only the verifiable time, speaker, affiliation, and title are retained, and no content inference is made.

75. AI 时代数据库生态的技术创新与实践 (Technological Innovation and Practice in the Database Ecosystem in the AI Era)

76. 开源多主数据库的突破 (Breakthroughs in open-source multi-master databases)

77. 从向量数据库到 AI 应用的数据底座 (From Vector Databases to the Data Foundation for AI Applications)

78. Vastbase: 智能驱动效率的 AI 数据底座 (Vastbase: AI Data Foundation for Intelligence-Driven Efficiency)

79. IntarkDB: 端侧智能的底层数据基础设施 (IntarkDB: The Underlying Data Infrastructure for Edge-Side Intelligence)

80. 圆桌: openGauss 7.0.0 LTS 技术亮点与数据库生态协同发展新趋势 (Roundtable: Exploring openGauss 7.0.0 LTS Technical Highlights and New Trends in Collaborative Database Ecosystem Development)

Agentic AI Infrastructure Forum (2F – KAI XUAN 2, afternoon)

For the 6 talks in this track, the official talk pages all mark "Abstract to be announced". The first 4 entries are Inferred based on public materials from the speakers' projects; for the last 2, no material on the same topic was found, so they are marked No material.

81. Running Millions of Agents at Tencent: CubeSandbox as the Hands Behind the Harness Brain

Details (inferred, sources below): Problem definition: when AI agents can dynamically write, compile, and run code, how to run this code safely becomes a core security issue — if an agent runs a malicious script or makes a mistake, it may access host files or bring down the entire server, while standard containers share the kernel and may not prevent escapes. Solution: Tencent Cloud CubeSandbox, written in Rust and fully open-sourced under Apache 2.0 on April 21, 2026, is positioned as a complete sandbox-as-a-service technology stack, not just an SDK. The four key features are, one by one: ① hardware-level isolation — based on KVM + RustVMM, each sandbox has its own independent Guest OS kernel (rather than a Docker-style shared kernel); a review article points out that it has its own hypervisor based on RustVMM, rather than wrapping Firecracker; ② cold start under 60 milliseconds (vendor data: average 67ms, P95 90ms at 50 concurrent), which is critical for real-time agents that need immediate code execution; ③ high density — the memory overhead of each sandbox instance is under 5MB, and a single physical machine can run thousands of fully isolated concurrent sandboxes; ④ drop-in replacement for E2B — fully API-compatible with the E2B SDK, and migration to self-hosting requires changing only one environment variable. Deployment: supports single-machine validation; the repository includes Terraform scripts for cluster deployment. Upcoming feature: snapshot and state-rollback will be open-sourced after it is fully completed. Adoption: reports say that MiniMax has used it to run hundreds of thousands of concurrent sandboxes. Caveat: 60ms cold start and similar figures are vendor data; one review notes that its performance outside Tencent's own bare-metal testbed still needs to be verified; the claim of being "the industry's only open-source agent sandbox" is a marketing statement. The above is inferred content, not what the speaker said at the conference.

82. vLLM and Ascend: 插件化架构与 MoE 专家负载均衡实践 (vLLM and Ascend: Plugin-Based Architecture and MoE Expert Load Balancing in Practice)

Details (Inferred; sources below): Architecture: vllm-ascend integrates into the vLLM upstream as a plugin. The EPLB module is located in the vllm_ascend/eplb directory and is split into two parts: the adaptor package (containing vllm_adaptor.py) and the core/policy package (containing policy_abstract.py, policy_default_eplb.py, and other policy classes). Problem to solve: under Expert Parallelism, expert load is uneven across NPUs. Mechanism: (1) replicate heavily loaded experts, then pack and place these replicas on each device; (2) an additional goal is to place experts in the same routing group on the same node to reduce cross-node traffic. Responsibility boundary: the policy code calculates replication and placement plans based on estimated expert load; load prediction itself is not in this repository, and the documentation mentions that a common practice is to apply a moving average to historical statistics. Extension method: to add a new balancing algorithm, register the policy type and class in PolicyFactory in policy_factory.py; to integrate a new MoE model, set model-related parameters such as the number of dense layers, the global number of experts, and the number of MoE layers during adaptor initialization, and register it through model_register; the reference case is Qwen3-MoE. Engineering recommendations: use decorators for non-intrusive patches, keep abstract class interfaces stable, and use Qwen3-MoE for benchmarking. The above is inferred content, based on official documentation rather than the speaker's statements.

83. From Rollout to Weight Synchronization: Mooncake 如何支撑高性能 RL 与推理 (From Rollout to Weight Synchronization: How Mooncake Supports High-Performance RL and Inference)

Details (Inferred; sources below): Mooncake Store provides fast weight update capabilities for RL and checkpoint scenarios by storing weights internally, and provides tensor-native and zero-copy APIs; Mooncake has joined the PyTorch ecosystem. In the broader pipeline, Checkpoint Engine is lightweight middleware for LLM inference engines that performs in-place model weight updates, provides broadcast and P2P update strategies, and Mooncake TE can serve as its pluggable transport backend; a paper (arXiv 2604.00368) used Checkpoint Engine v0.2.0 + Mooncake TE as a baseline on 8×H800 to compare weight update times for Qwen3-235B and GLM-4.5-Air. The "Rollout → Weight Synchronization" in the title corresponds exactly to this pipeline in RL post-training: after sampling generates trajectories, the updated weights need to be synchronized back to inference replicas quickly. The above is inferred content, based on third-party documentation and papers, not the speaker's statements.

84. Triton For Ascend: 简化高性能算子开发 (Triton For Ascend: Simplifying High-Performance Operator Development)

Details (Inferred; sources below): Ascend's clear strategy for ecosystem integration is "not to make developers passively adapt to Ascend, but to make Ascend actively adapt to developers' habits." Triton Ascend is a flagship result of this strategy—it became the first domestic accelerator backend supported by the official Triton community. The same batch of results also includes TorchNPU launching on the official PyTorch website (the first Chinese chip hardware to receive official support, and also the first accelerator backend with official branch repository download support). Ascend as a whole has been adapted to more than 90 mainstream open-source projects, including PyTorch, vLLM, SGLang, Megatron, Triton, and TileLang. "Simplifying High-Performance Operator Development" in the title means that developers can use Triton's programming model to produce high-performance operators on Ascend. The above is inferred content.

85. AscendNPU IR: 构建高效易用的昇腾多样性算子体系, 赋能开发者灵活创新 (AscendNPU IR: Building an Efficient and Easy-to-Use Ascend Diverse Operator System, Empowering Developers to Innovate Flexibly)

86. KunPeng SuperPoD - 面向 Agentic AI 的新型算力基础设施 (KunPeng SuperPoD - New Computing Infrastructure for the Agentic AI)

Spotlight Shenzhen 2026: Next-Gen AI-Native Products (3F – ADEN 1)

87. Spotlight Day 1 Roadshow

Details (official abstract): Spotlight Shenzhen 2026 roadshow Day 1. The first 5 minutes introduce Spotlight itself, followed by demonstrations of 8 projects; all times are China Standard Time (UTC+8). The official minute-by-minute schedule is as follows:

Time Project
12:30–12:35 Spotlight introduction
12:35–12:42 Petoi Quaddle
12:42–12:49 Fugleramme
12:49–12:56 Hen Local Translator
12:56–13:03 Redbank III
13:03–13:10 AI Desk Phone
13:10–13:17 reset
13:17–13:24 Xiapan USB
13:24–13:30 PhyAgentOS

The official schedule does not provide introductory text for each project, so it is not possible to write a key argument for each project. (Note: PhyAgentOS also has a full talk on Saturday in the Open Source Robotics track, titled "PhyAgentOS: Recursively Self-Evolving Physical Agent Operating System," with speaker Tao Pu.)

Appendix: Statistics by Source Type

Source type Count Notes
Abstract only (with substantive official abstract) 70 This document expands with a complete translated account based on the abstract
Inferred (no official abstract; supplemented with same-topic material from the speaker/project) 8 Inferred paragraphs in items 2, 3, 35, 71, 72, 73, 74, 81, 82, 83, and 84 (item 1 has dual sources: "Abstract + Inferred")
No material (no official abstract and no same-topic material) 9 Items 65, 68, 75, 76, 77, 78, 79, 80, 85, and 86

Note: Item 1 has both an official abstract and an inferred supplement; items 2, 3, and 35 are "no/weak official abstract + inferred"; for statistics, they are classified by primary source, so the sum of the columns is not exactly equal to 87.

Raw data files:

GOSIM Shenzhen 2026 · Saturday (October 17, Day 2) Detailed Talk Notes

Overview

Dates and venue. GOSIM Shenzhen 2026 will be held from October 16, 2026 (Friday) to October 17 (Saturday) at ADEN Hotel Shenzhen Nanshan. This document covers only Saturday. For the agenda, the official website is authoritative: https://shenzhen2026.gosim.org/schedule/. The official link for each talk in this document is https://shenzhen2026.gosim.org plus the path given for that entry.

Source limitations. The conference ended only yesterday. As of October 18, 2026, no post-conference coverage, recordings, or speaker slides have been found. The CSDN, Tencent News, and Zhihu articles found are all pre-conference previews. Therefore, the content of this document has only two sources:

Count. The official website lists a total of 84 agenda entries for Saturday, excluding lunch, tea breaks, and check-in. Of these, 78 are substantive talks, and 6 are procedural items (remarks, AMA, roadshow, awards ceremony, closing ceremony). There is also a "Rust Instructor Training & Workshop" (2F HAI HAN), invitation only, time TBA; the official website states that its content will not be made public, and it is not included in the count.

Track Venue Count
Keynote Plenary 3F ADEN full hall 3
Agentic AI Summit 3F ADEN 1 12 (including 2 procedural items)
Agentic AI on Edge 3F MING HUA (morning) 4
Agentic OS & App 2F HAI XIA 2 9
Agentic Device 2F KAI XUAN 1 8
Open Source Models & Infra 3F ADEN 2 9
Open Source Robotics 2F HAI XIA 1 10
KVCDN Workshop 2F KAI XUAN 2 9
vLLM Workshop 3F MING HUA (afternoon) 8 (including 2 procedural items)
AI Education Workshop 3F HAI LIAO (afternoon) 8
Google Cloud Workshop 3F HAI LIAO (morning) 2
Spotlight Shenzhen 2026 3F ADEN 1 2 (both procedural items)
Total 84

Title-link mismatch. For several talks, the official link slug does not match the title; for example, OctoLoop's link slug is hafleet, and Shaomin Zhu's link slug is agentos. The official website appears to have changed the titles but not the links. This document uses the titles and abstracts displayed on the page as the authority.

I. Keynote Plenary (Main Venue, 3F ADEN)

1. Omarchy: Beautiful, Fun & Agentic Linux

2. 从开源软件到开源智能: 硅基时间正在重写开发者世界 (From Open Source Software to Open Intelligence: Silicon Time Is Rewriting the Developer World)

3. Multimodel Intension Model

II. Agentic AI Summit (3F ADEN 1)

4. Mapping the Open Agentic AI Landscape: From Open Models to Cross-Foundation Architectures

5. Next-Generation System Interoperability: AI Agents as the Universal Integration Layer

6. Fireside Chat

7. Evaluation You Can Audit: Eclipse PanEval as Shared Infrastructure

8. Latest LF Research Reports on Agentic AI and AI Coding

9. Network-Sourced AI: Open Ensembles, Private Data, and a Reproducible Path to the Frontier

10. Cangjie and the New Frontier of Programming Languages in the AI Era

11. The Agentic AI Developer Economy: Building a Meritocratic Open Source Ecosystem

12. Your AI Product Is Ready. Is Your Company Ready for Europe?

13. The Sovereignty Layer: Why Open Source Is How the Global South Shapes AI

14. Hackathon Summary and Awards Ceremony (program item)

17:30–17:50, Yun Lin (Associate Professor, Shanghai Jiao Tong University) and Zonghuan Wu (MoFA.AI) present the hackathon summary and awards. /schedule/summit-hackathon-awards/

15. Closing Ceremony (program item)

17:55–18:00, conference closing ceremony. /schedule/summit-closing-ceremony/

III. Agentic AI on Edge (3F MING HUA, morning only, ends at 12:15)

16. Agentic AI on Global Edge: Deploying SPEAR-Powered Operations Agents with MCP

17. DHttp: From Partial Connectivity to Omniconnectivity

18. From Khronos AI Ecosystem Research Insights to Edge AI Inference with Vulkan

19. Which operating system is best for agentic mobile devices?

IV. Agentic OS & App (2F HAI XIA 2)

20. Agent Harness for Multiplayer Collaboration

21. OctoLoop: Making Agents Truly Loop — Inner-Loop Autonomy, Outer-Loop Correction, and Evolution-Loop Experience Accumulation

22. MemOS: The Next-Generation Memory System for Autonomous Agent Evolution

23. Introducing Harness Intelligence Model: JIT-Agent

24. Voice AI Is Entering Its Second Generation: Lessons from Enterprise Voice Agent Deployments

25. Agent Harness for Server Firmware Development

26. An Autonomous Runtime Shielding Agent for Linux Kernel Vulnerabilities

27. Jiuwen Agent Memory System

28. From Coding Agents to Autonomous Contributors: Designing the Operating Model for Agentic Software Engineering

5. Agentic Device (2F KAI XUAN 1)

29. 开发者的 AI 创造: 技术社区视角下共建 AI 硬件 (Developers' AI Creations: Co-building AI Hardware through the Lens of a Tech Community)

30. The World as the Interface: Building an Open Agent Ecosystem for AI Glasses

31. Building AI Agent Devices on RISC-V with DC-ROMA Mainboard III

32. AI 原生嵌入式产品研发: 重塑软硬件研发范式 (AI-Native Embedded Product Development: Reshaping the Hardware-Software R&D Paradigm)

33. Building the Infrastructure for the AI World

34. You Built the Prototype. Should It Become a Product?

35. How Hiwonder Helps Students Apply AI Agents on Robots

36. AI 雷达: 感知加计算带来的行业变革 (AI Radar — Industry Transformation Driven by Sensing + Computing)

VI. Open Source Models & Infra (3F ADEN 2)

37. Teaching AI to Improve AI: From Experimental Feedback to Meta-Evolution

38. Towards Interaction Intelligence

39. 从开源模型到开发者生产力: AI 时代的开源基础设施重构 (From Open-Source Models to Developer Productivity: Reconstructing Open-Source Infrastructure in the AI Era)

40. TransferQueue: An Asynchronous, High-Performance Streaming Data Engine for Post-Training

41. Efficient Training of Large Language Models via Orthogonal Transformation

42. C-to-Rust Code Translation: A Three-Stage Curriculum of Pretraining, Debugging-Aware SFT, and Task-Specific SFT

43. Building Self-Improving and Long-Horizon Agents

44. AI-Ready Data Production and Dynamic Data Training for Open-Source Models

45. Applications of Voice Agents in Embodied Intelligence and High-Quality Data-Driven Solutions

7. Open Source Robotics (2F HAI XIA 1)

46. "World Model for All": Towards Unified Generation, Navigation and Control

47. 具身智能 VLA 算法 (Embodied Intelligence VLA Algorithms)

48. Everyone Can Build an Open-Source Robot

49. From Leaderboards to Global Infrastructure: Building an Open Evaluation Ecosystem for Embodied AI

50. From Single-Robot Intelligence to Multi-Agent Coordination: The Evolution of AIRSEAI

51. Rethinking Efficiency Bottleneck of Generative Robotics Manipulation

52. Interactive Heuristic Learning Method of LiDAR and Camera for Target Detection

53. AMD ROCm and Radeon-Powered Real2Sim2Real Vision-Language-Action Manipulation

54. PhyAgentOS: 递归自进化物理智能体操作系统 (PhyAgentOS: Recursively Self-Evolving Physical Agent Operating System)

55. Full-Stack Embodied AI OS — Introduction to openEuler Intelligence Boom Robot (IB-Robot)

8. KVCDN Workshop (2F KAI XUAN 2)

56. High-Quality Token Service: From Heterogeneous PD to Cross-Region AI Infrastructure

57. LMCache's Production-Grade Multi-Process KV Cache: Architecture, Observability & Fault Tolerance

58. KV Cache Usage in SGLang

59. UCM: Beyond KVCache Acceleration, Lifecycle Management for Agent-Scale Inference

60. 聊聊运用 FlexKV 进行大模型 KVCache 池化的最佳实践 (Best Practices of Pooling LLM KVCache with FlexKV)

61. Research and Practice of Agent Networking

62. Turbo: Efficiently Serving Long-Context LLMs with In-Network Aggregation

63. Beyond Compression Ratio: Service-Aware KV Cache Compression for Distributed LLM Serving

64. Mooncake Store for Agentic RL in Disaggregated Architectures

IX. vLLM Workshop (3F MING HUA, afternoon; check-in at 13:10, starts at 13:40)

65. Opening Remarks (Procedural item)

13:40–13:50, Hongsheng Liu (vLLM-Omni core maintainer and chief architect of the InferMatrix project) and Chuhui Chen (Red Hat) deliver opening remarks. /schedule/vllm-opening-remarks/

66. vLLM Latest Update and Roadmap

67. llm-d: 打造跨负载,跨模态,跨硬件,跨平台的分布式推理平面 (llm-d: Building a Cross-Workload, Cross-Modal, Cross-Hardware, and Cross-Platform Distributed Inference Control Plane)

68. AFD Plugin: Attention–FFN Disaggregation in vLLM

69. 从单点加速到系统最优: 大模型推理的全栈优化实践 (From Single-Point Acceleration to System-Wide Optimization: Full-Stack Optimization Practices for LLM Inference)

70. Inside ROCm 10.0: Building an Open GPU Software Platform for the Agentic AI Era

71. Scaling vLLM System-Wide: From Network I/O to Distributed Scheduling

72. AMA with the vLLM Core Team (Program Item)

17:10–17:30, public Q&A with the vLLM core team./schedule/vllm-core-team-ama/

10. AI Education Workshop (3F HAI LIAO, Afternoon)

73. From Passive Lectures to Active Co-Learning: The AI-Interactive Future Classroom

74. 具身智能领域人才需求 (Talent Demand in the Field of Embodied AI)

75. 开源机器人与 AI 教育的全学段贯通实践 (Full-stage Integrated Practice of Open-source Robots and AI Education)

76. Software Engineering 3.0: Exploration and Practice of Innovative Talent Cultivation Based on Open Source and AI

77. 对话即创造: 用 AI 智能体重构中小学科技教育 (Dialogue Is Creation: Reshaping Primary and Secondary School STEM Education with AI Agents)

78. 当孔子遇见维果茨基: AI 赋能时代的学习重构与未来学校建设 (When Confucius Meets Vygotsky — Learning Restructuring and Future School Construction in the AI-Empowered Era)

79. From AI Access to AI Impact: Building Inclusive Education Across Cultures

80. Exploration and Practice of New Global Models for Industry-Academia-Research Education in the AI Era (Roundtable)

11. Google Cloud Workshop (3F HAI LIAO, morning, hands-on session, bring your own laptop)

81. Get started with Agent Development Kit (ADK)

82. Build multi-agent systems with ADK

12. Spotlight Shenzhen 2026: Next-Gen AI-Native Products (3F ADEN 1)

83. Spotlight Day 2 Roadshow (Program item)

12:30–13:30. 12:30–12:38 VACAT AI Visual Creativity Awards ceremony; 12:38–12:43 Spotlight introduction; followed by six project pitches of about 8 minutes each, in order: VibeKeys, Mira Light, Octos Learn, Meiso Glass, Ingora · Qrio, PoopSense. The official website does not provide descriptions of each project./schedule/spotlight-day-2-roadshow/

84. Spotlight Shenzhen 2026 Awards Ceremony (Program item)

17:50–17:55, Spotlight awards presentation./schedule/summit-spotlight-awards/