ParisMay 19, 2026 /PRNewswire/ — As large models enter the engineering development phase, the industry’s focus is shifting from mere model capability breakthroughs to achieving stable deployment, low-cost operation, and long-term efficient performance. Meanwhile, agent frameworks represented by OpenClaw are gaining increasing traction, with AI’s ability to autonomously execute various tasks becoming ever closer to practical production scenarios. However, real-world challenges such as high inference costs, complex invocation chain structures, and the difficulty of sustaining long-term open-source ecosystem development have emerged as new hurdles the entire industry must urgently overcome.
Against this backdrop, GOSIM Paris 2026 was officially held in Paris from May 4 to 6.
This open-source AI technology conference, aimed at global developers, was organized by GOSIM and co-created by CSDN, 1ms.ai, and Probabl. The focus of the discussions was no longer limited to exploring the performance advantages of large models themselves, but rather on the practical application and deployment of AI technology capabilities, with a strong push to deeply integrate related technologies into engineering projects, complete toolchains, and various real-world business scenarios.
The conference featured a rich array of activities including keynote speeches, themed forums, hands-on workshops, and hackathons. On-site attendees included engineers, researchers, and industry practitioners from renowned institutions such as Probabl, Huawei, NVIDIA, ByteDance, Hugging Face, Eclipse Foundation, Amazon, Fundamental, LAION, Beijing Academy of Artificial Intelligence (BAAI), Portland State University, Neo4j AI Community, and Fraunhofer IAIS. On one hand, technological capabilities are advancing rapidly; on the other, there is continuous debate and collaboration around open-source ecosystems, engineering collaboration, and practical implementation.
GOSIM Paris 2026 Conference Scene
As model capabilities gradually become homogenized, the true differentiator is no longer just parameter scale, but whether the engineering system is robust, whether the ecosystem is sustainable, and whether collaboration efficiency can truly support AI’s entry into long-term production environments.
This is precisely the question GOSIM Paris 2026 sought to answer.
Focusing on AI’s Origins and Industrial Transformation: Keynotes Take Center Stage
During the keynote session at the main venue of GOSIM Paris 2026, Sir Timothy Gowers, Fields Medalist (the “Nobel Prize” of mathematics), Professor of Combinatorics at the Collège de France, Research Professor at the University of Cambridge, and Fellow of Trinity College, Cambridge, delivered a keynote speech titled “Can an AI Mathematician Be More Than a Black Box?” From the perspective of a frontline mathematics researcher, he systematically outlined the impact and potential trajectory of large language models on mathematical research methods. Sir Timothy Gowers emphasized that the current form of AI mathematicians is still not ideal. Compared to human mathematicians, large models often provide direct answers but lack the ability to guide the process. In learning and research, humans need step-by-step prompts and a “productive thought process” rather than just the final conclusion.
Following this, Gaël Varoquaux, Chief Scientific Officer at Probabl, steered the conversation towards a more “fundamental” issue: despite the remarkable achievements of large language models in image generation, text dialogue, and even theorem proving, seemingly far removed from noise and randomness, statistical tools remain the underlying logic supporting these accomplishments.
Bill Ren (Ren Xudong), Huawei Open Source Liaison Officer and CNCF Board Member, stated that code has become the new language, and the boundaries of open source will define the boundaries of the future world. “The next constitution of civilization will be code – and code must be open. The assumptions of the Carbon Era are outdated; the first defining features of the Silicon Era have yet to be written. We are jointly building a shared destiny digital community – led by artificial intelligence, arbitrated by humans, governed globally, and built upon an open foundation.”
Jiang Tao, Founder and Chairman of CSDN and Initiator of GOSIM, provided a systematic summary of this round of AI transformation from a broader industrial evolution perspective. Jiang Tao pointed out that with the significant reduction in the cost of intelligent technology, the era of programmable everything has officially arrived. Looking at the entire agent technology system, open-source capabilities already cover all levels. The current core competitive track lies in the coordination and scheduling layer. As unified industry standards are still in their infancy, whoever can establish the rules first will dominate the development of the next era.
Subsequently, Niko Matsakis, Senior Principal Engineer at Amazon, turned his attention to a forward-looking direction: how to design the Rust toolchain from the outset to be suitable for the age of agents. His open-source project, Symposium, is a practical exploration aimed at this goal.
Against the backdrop of large model performance approaching bottlenecks, the focus of competition in the AI industry is shifting from “model capability” to “system capability.” Yonghua Lin, Vice President and Chief Engineer at the Beijing Academy of Artificial Intelligence (BAAI), shared insights on the theme “From ChatGPT to OpenClaw: How Agentic AI Redefines AI Computing and Open Evaluation,” pointing out that infrastructure is becoming the next critical bottleneck constraining AI development.
Alexandre Gerbeaux, Head of Applied AI at Fundamental, drew from his personal career experience to outline the development path of data science and introduced an emerging new direction – Large Table Models (LTM). He used a vivid metaphor to describe the fundamental relationship between large models and table models: if large language models are more like AI’s “right brain,” excelling at generation and creation, then table models are closer to the “left brain,” responsible for logic and numerical reasoning.
GOSIM Paris 2026 AI Vision Forum Kicks Off
As one of the flagship events of GOSIM Paris 2026, the GOSIM Paris 2026 AI Vision Forum brought together invited guests from around the world for candid discussions under the Chatham House Rule, exploring a core proposition concerning the future of humanity: as AI reshapes human roles, how can we ensure that human core values remain central?
Guided by host Jesse McCrosky (Principal AI Architect, Egen AI), Yann Lechelle, Executive President and Chairman of Probabl, drew on his four decades of software experience to paint a grand picture of time, technology, and the human role for the audience. He shared the seven pillars of open computing infrastructure – Open Science, Open Data, Open Standards, Open Source Software, Open Weights, Open Platform, and Open Hardware – and stated: “Without these seven pillars, there is no real AI. This is the answer we must develop, we must promote, we must demand transparency for – only then will the algorithms not devour us for breakfast tomorrow.”
Sir William Timothy Gowers, Fields Medalist, Professor of Combinatorics at the Collège de France, Research Professor at the University of Cambridge, and Fellow of Trinity College, Cambridge, subsequently noted: “There is an anxiety within the mathematical community – that we may be entering a world where those with access to powerful models will be able to prove great results using them, while those without access will be at a tremendous disadvantage.”
The forum also featured four special roundtable discussions covering the following core topics:
- Agentic AI Systems – Several on-site guests quickly delved into different angles: some emphasized that open computing, open data, and open evaluation will become the infrastructure for the agent era; others pointed out that existing open-source licenses and standard systems face new challenges given the uncertainty of agent behavior; regarding software engineering practices, some believed engineers are shifting from “writing code” to “orchestrating systems”; concerning talent development, it was suggested that the growth path for junior developers needs redesigning in an environment where agents participate in development; and regarding model selection, the practical differences between open-weight and closed-source models became a recurring point of comparison.
- AI in Education Applications – This discussion began with a themed presentation. On-site guests explored from a “cognitive design” perspective why 95% of current AI education pilot projects fail and how agents can truly empower learning through better cognitive architectures.
- Trustworthy AI Governance – During this roundtable, several core viewpoints emerged: existing regulatory frameworks must adapt to the dynamism and uncertainty of agents, while transparency, explainability, and fairness remain the unshakeable foundations of trustworthy AI. Turning to multi-agent systems, interoperability standards were seen as key to scaling; correspondingly, identity verification and cross-system collaboration urgently require clearer international rules. Looking deeper, cryptography-based verification systems are becoming critical infrastructure supporting sovereign AI.
- Open Source and Digital Public Goods – This session brought together representatives from system-level software foundations, leading large model companies, international research institutions, and developer communities to discuss the infrastructure foundation for the Agentic AI era.
Group Photo of GOSIM Paris 2026 AI Vision Forum Guests
Unlocking the Full Ecosystem Evolution Path of Agentic AI
Alongside the keynote sessions, GOSIM Paris 2026 simultaneously hosted five vertical themed forums, covering five core directions: Agentic AI Summit, Open Source Models, Agentic OS and Applications, Agentic AI on Edge, and Open Source Robotics. From foundational model R&D and system application building to edge intelligence deployment and open-source robotics innovation, these forums focused on real industry needs, deconstructed technical implementation challenges layer by layer, and explored the practical value of open source and agent technology through in-depth exchanges within specific domains.
At the “Agentic AI Summit” sub-forum, the value of Agentic AI was no longer confined to the conceptual level but was directly validated and expanded within real industrial scenarios. This Agentic AI Summit brought together leading retail enterprises, representatives of domestic foundational software, and university researchers, systematically presenting the practical pathways and application boundaries of Agentic AI across different industry dimensions, from enterprise agent deployment and R&D process efficiency to the intelligent transformation of education systems.
Simultaneously, at the “Open Source Models” forum, frontline technical experts from institutions including Moonshot AI, Zhipu AI, Minimax, Northeastern University, the Institute of Automation of the Chinese Academy of Sciences, BAAI, LF AI & Data Foundation, Huawei, and Shanghai Jiao Tong University shared the latest progress and practical experiences in areas such as training methods for open-source large models, inference optimization, multimodal capabilities, community collaboration, and industrial deployment practices.
At the “Agentic OS and Applications” forum, frontline experts from companies like Makepad, Delinea, CopilotKit, Huansheng Technology, Huawei, ByteDance, Makeitfuture, FOSS Shanghai, Eclipse Foundation, and Typeform discussed Agentic OS. Attendees generally agreed that as AI agents gradually become the primary interface for human-computer interaction, the role of the operating system is also changing – it is no longer just a traditional collection of kernels and drivers but is evolving into the infrastructure supporting autonomous, composable, and trustworthy AI workflows.
When large models no longer rely entirely on cloud computing power, and intelligent terminals can operate independently in offline and low-power environments, Agentic AI on Edge is reshaping how AI is deployed. The conference specifically featured the “Agentic AI on Edge” forum, discussing the full technology chain and practical pathways for edge-side AI. Technical experts from Hugging Face, NVIDIA, Dimforge, Second State, and other institutions provided systematic insights ranging from frameworks to engineering implementation.
Beyond the “text world” of LLMs, robots are becoming a key vehicle for AI to connect with the physical world. However, for a long time, high hardware costs, fragmented software environments, and inconsistent evaluation standards have kept “embodied intelligence” mostly confined to demo videos, making large-scale reproduction and production deployment difficult. At this open-source robotics forum, a new possibility emerged: from real-robot evaluation standards and humanoid robot platforms to model self-evolution and bionic dexterous hands, an open-source-driven embodied intelligence technology stack is taking shape.
When AI Discussions Turn to Practice, On-Site Workshops Let You Get Hands-On!
In addition to high-level keynote speeches and specialized forum discussions, this year’s conference also emphasized hands-on practice for developers. Workshops including vLLM Workshop, OpenHarmony × AI: Powering the Next Intelligent Operating System, Open Computing for Agentic AI (FlagOS), Data Science and AI Workshop, SGLang Workshop, and Building Agentic Applications Workshop took place in succession. These sessions focused on technical practice, underlying system architecture, open agent ecosystems, and core data science methodologies, providing developers and industry practitioners with immersive learning and practical exchange through hands-on teaching, case analysis, and collaborative creation.
At the vLLM Workshop, Erwan Gallen, a core contributor to vLLM and Senior Principal Product Manager for Generative AI at Red Hat, deconstructed the key trade-offs in infrastructure selection from a platform team’s perspective, covering different workload patterns, differences in prompt and context behavior, scaling architecture patterns, and operational adaptation issues in production environments. Subsequently, Daniele Trifirò, Principal Software Engineer at Red Hat, started with the core optimization mechanisms of vLLM, focusing on key designs including PagedAttention, and further extended to system construction methods for different target scenarios.
At the “OpenHarmony × AI: Powering the Next Intelligent Operating System” Workshop, multiple guests from open-source communities, academia, and industry shared their latest practices and challenges regarding the integration of intelligent operating systems and AI.
As agent systems rapidly move towards engineering implementation, how to build a scalable, evaluable, and governable open computing foundation became the core discussion point of the “Open Computing for Agentic AI (FlagOS)” Workshop. Representatives from BAAI, the Eclipse Foundation, members of the ICCSD UNESCO Advisory Committee, the LF AI & Data Foundation, and the SGLang community discussed open ecosystems, evaluation systems, and global collaboration pathways.
As large models move from the lab to the production line, how can the gap between data science and AI engineering be bridged? At the “Data Science and AI Workshop,” several frontline experts from renowned European institutions and open-source communities provided practical answers from their respective fields, covering foundational model building, large-scale training, synthetic data, efficient OCR, and agent evaluation.
The “SGLang Workshop” was a hardcore practical session for developers. Xinyuan Tong, an open-source maintainer of SGLang, deconstructed the design principles behind SGLang’s performance. Eva Ma, Algorithm Engineer at Atlas Cloud AI LLC, and Yuhao Yang, a SGLang developer, demonstrated from both presentation and practical angles how SGLang-Diffusion provides production-grade inference capabilities for image/video generation through advanced parallelization, distributed VAE, and operator fusion. Finally, Ethan (Yusheng) Su, a technical team member at RadixArk, directly showcased SGLang’s core role in RL post-training, using the Miles RL framework with SGLang as the sampling backend to build an end-to-end RL training pipeline.
At the “Building Agentic Applications Workshop,” the discussion focused on how to turn AI capabilities into truly runnable engineering systems, rather than staying at the conceptual level. Rik Arends, Co-founder of Makepad, and Bart Massey, Associate Professor of Computer Science at Portland State University, provided hands-on guidance and live instruction.
“What comes after LLMs?” This question, of great interest to many, also became one of the discussion topics at the GOSIM Paris 2026 lunchtime learning session. Alexandre Gerbeaux, Head of Applied AI at Fundamental, pointed out in his presentation that Large Table Models (LTMs) are emerging as a new model category specifically designed to process and understand structured data, thereby bridging the gap between current model capabilities and enterprise data realities.
Geeks Compete on Stage, Hardcore Works Unveiled
The innovation competition arena saw its highlight moments. GOSIM Spotlight 2026 Frontier Creators Ceremony, the dual-track hackathons “Agentic Hackathon” and “Robotics Hackathon,” and the “FlagOS KernelGen 24-Hour Bounty Challenge” were successfully concluded. After two days of intense competition from May 5 to 6, various innovative projects and practical outcomes were showcased, providing a vibrant conclusion to this Paris conference.
On the GOSIM Spotlight 2026 Frontier Creators track, covering areas from multimodal generation and interactive expression to toolchains and content format experiments, 10 finalist creators presented their works and shared insights at Station F. These projects were not just about “using tools to create content”; they were more about solving specific problems, with each project addressing a clear creative pain point.
On the hackathon front, two events ran concurrently. The Agentic Hackathon covered multiple tracks including text, voice, music, video, and presentation generation. Participating teams rapidly built application prototypes using models like Zhipu GLM, Moonshot Kimi, and MiniMax, transforming ideas into runnable demonstrations within two days. The Robotics Hackathon was more “hands-on,” with all tasks completed on the fully open-source OpenArm humanoid robotic arm, including grasping and assembling, liquid pouring, fabric folding, and human-robot handovers. Participants continuously debugged mechanical structures and control codes on-site, spending almost the entire time solving problems practically.
The FlagOS KernelGen 24-Hour Bounty Challenge was equally exciting. Centered around the Triton language, this challenge focused on kernel code generation and optimization, placing high demands on participants’ engineering skills and tool usage efficiency.
From content creation to robot operation, from application-layer prototypes to system-level optimization, these award-winning projects shared a common trait: they could all run on-site and provide answers to specific problems.
Award Ceremony at GOSIM Paris 2026
GOSIM Paris 2026 Witnesses the Thriving Open-Source AI Ecosystem, Looking Forward to Reuniting in Hangzhou in October!
With this, GOSIM Paris 2026 officially concluded. As an event dedicated to global open-source collaboration, this Paris gathering brought together developers and industry practitioners from around the world, fostering lively exchanges and continuous intellectual碰撞 (collision of ideas).
Of course, the smooth execution of the entire conference was made possible by the concerted efforts and collaborative partnership of various enterprises, open-source communities, and ecosystem partners.
We extend our heartfelt gratitude to our premier sponsor, Huawei, as well as partners including OpenHarmony, Fundamental, OpenBMB, MiniMax, KIMI, Zhipu AI, Novita, RadixArk, and Databricks for their significant support and generous assistance. We also sincerely thank all supporting partners such as BAAI, OuiCrea, the All-China Youth Innovation and Entrepreneurship Association in France, the Sino-French Artificial Intelligence Association, Olares, Second State, the SGLang community, the vLLM community, the FlagOS community, Hugging Face, Kaiyuanshe, Unaite, 42-ai, AI By The Bay, InnAIO, and Jumeau.AI for their enthusiastic participation and active involvement.
Various exhibitors and ecosystem partners gathered on-site, contributing to industrial synergy through diverse displays and collaborative creation, continuously enriching the open-source AI ecosystem landscape and adding vibrant energy and industrial value to this international event.
Three days of intellectual exchange and practical exploration have concluded successfully, but the journey of cross-regional open-source collaboration continues. This event built an efficient global dialogue bridge, accumulating a wealth of practical experience and industry consensus. We look forward to reuniting in the future, continuing to explore new directions in intelligent technology together, and building a more open and prosperous global open-source ecosystem.