{"slug": "gosim-paris-this-is-what-open-source-ai-looks-like-in-2026", "title": "GOSIM Paris: This Is What Open Source AI Looks Like in 2026", "summary": "GOSIM Paris 2025, held at Station F on May 5-6, showcased open-source AI developments including LLMs advancing in mathematical reasoning, a call for transparency over speed, and the introduction of TabICL for tabular data. The conference featured technical workshops on vLLM, highlighting PagedAttention and multi-hardware support, emphasizing the importance of open-source tools in AI.", "body_md": "GOSIM (Global Open Source Innovation Meetup) is an open-source conference whose latest edition took place in Paris on May 5–6 at Station F. It brought together the people who actually build the open-source tools we all rely on. It’s one of those conferences that doesn’t get much attention, but probably deserves more.\n\nA few takeaways from a day spent exploring open-source AI.\n\nLLMs are getting good at math, fast. I was curious to hear what a Fields Medal mathematician actually thought about it. Sir Gowers made a point: a correct proof isn’t enough. He mentioned PhD students feeling overwhelmed by how quickly AI solves problems they’ve spent months on (and the illustration was insane lol), and then an LLM solves it in a matter of minutes. His group at Cambridge is actually working on making AI reasoning transparent and teachable, not just fast. The shift from *solve it* to *show me how you got there* feels like the more interesting race right now.\n\nWe were in the middle of a conf dominated by agentic AI, and then Gaël Varoquaux (Chief Science Officer at Probabl) walked on stage and made a simple point: you don’t need an LLM for everything. Data science is reinventing itself, and statistics still matter. He introduced TabICL, an open source foundation model for tabular data that holds its own against the best. Probabl, the team behind scikit-learn, is proving that “old school” open source data science is evolving without losing what made it work. In a room full of agent frameworks, that felt like a necessary reality check.\n\nWe kicked things off with these two keynotes, which I found particularly relevant. But I’m not going to list every talk one by one here, you can already find the slides online. What I really wanted to do instead was share my thoughts on the conference as a whole. I’ve become a bit of a Paris tech conference enthusiast, attending events that range from deeply technical to much broader ones (I’ll let you guess which are my favourites).\n\nAfter this dose of big picture, it was time for the workshops, and I had a plan.\n\nRight after the two keynotes, I headed to the technical workshops. I had already planned my day around one session: vLLM. For a bit of context, it’s a library we’ve been using at work for inference for a while now, and the idea behind it is still genuinely fascinating (I’m currently writing an article on LLM inference and fine-tuning, so stay tuned. Yes, this is a shameless plug 🫣).\n\nDuring the workshop, Daniele Trifiro, vLLM contributor and Principal Software Engineer at Red Hat, started by presenting the core idea behind vLLM: PagedAttention, a KV-cache memory management approach inspired by OS paging, designed to drastically reduce memory waste and unlock fast inference. He covered recent optimizations, and I was genuinely surprised by how many new features I hadn’t heard of, and especially by the number of parameters I didn’t know existed, despite having used vLLM regularly.\n\nThe most interesting part was definitely the live code walkthrough , yes, literally. In big conferences (Not talking about conf like ICML, NeurIPS and so on), workshops are often theoretical, short, and don’t go very deep. This was the complete opposite: we had time to explore the library, and to dive into the hardware side, including vLLM’s support for Google TPU, AWS Neuron, and Intel Gaudi (not just NVIDIA GPUs).\n\nHe also covered evaluation, a crucial topic in this space, using [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) to check that vLLM actually meets requirements. The talk was a good opportunity to discuss the community, the importance of open-source contributors, and leaderboards like OpenLLM.\n\nI think you get the picture, I wasn’t disappointed. It was exactly the kind of technical deep dive I was hoping for: over an hour and a half of genuinely interesting content that captured the full scope of work behind vLLM.\n\nIf vLLM is about running a model efficiently, FlagOS tackles a more upstream question: how to run it everywhere. But before getting into that, a quick word on the agentic AI track in general.\n\nWhat would a 2026 conference be without agentic AI? There was something for everyone, from workshops to panel discussions.\n\nIf you’re starting to know me (or not), panels are not really my thing, and this time I sometimes felt a gap between the advertised topic and what was actually being discussed. You’d walk into a room expecting a conversation about agentic AI and end up hearing about strict European AI regulations. Interesting topic, of course, just not what you came for.\n\nOn the workshop side **FlagOS** tackles the challenge of running models across multiple hardware targets with a unified stack. The BAAI team, led by Yonghua Lin (Vice President & Chief Engineer at BAAI), explained how they built the framework, with FlagGems, FlagTree (multi-target compiler), and FlagScale (a non-invasive plugin for frameworks).\n\nWhat struck me most was the circularity of FlagOS. The loop is real: FlagOS runs the agents, and the agents build FlagOS in return. Concretely, complex infrastructure tasks, like 13-step model backporting with end-to-end precision validation, or kernel generation via MCP, are now handled autonomously by agents, with 99% correctness on multiple operator benchmarks. The infrastructure isn’t just agent-ready anymore, it’s agent-built and continuously improved by agents. Fascinating!\n\nThey then shifted the focus to evaluation via **FlagEval** and **PanEval, **a rethought approach to benchmarking (contributed to the Eclipse Foundation). The project is already available [ here](https://projects.eclipse.org/proposals/eclipse-paneval), though still fairly recent.\n\nA complete change of pace here, stepping away from infra and back to the fundamentals of data science.\n\nOne session that caught my attention in the Probabl track focused on the Skore library. I was already loosely familiar with it, but this talk made me look at it differently.\n\nSkore is built for enterprise data science: not the solo notebook, but teams, organizations, and scale. The idea is to make tracking, exploring, and sharing workflows as simple as a fit() / predict().\n\nThe session itself was short but dense. What stood out wasn’t the discovery of the library, but the way it was framed. I already knew the basics, but Marie Sacksick, Product Engineer, and Fabien Pesquerel, Developer Relations Engineer at Probabl, surfaced features I hadn’t really paid attention to before, especially around collaboration and workflow tracking.\n\nI left with a very concrete idea: turning it into a team presentation during one of our tech watch sessions at work. In just 20 minutes of live coding, it was enough to make me want to dive back into the docs, and potentially contribute.\n\nFor context, we regularly run open-source Saturdays, often working on Probabl’s Scrub library. Last weekend, I even found myself going through the Skore docs again. That’s usually the sign a conference really did its job.\n\nTo close the day on a high note, something a little different, and a topic I care deeply about.\n\nAs a volunteer with the Paris chapter of Women in Machine Learning & Data Science (WiMLDS), I was particularly happy to see these topics included in such a technical conference. We spend a lot of time talking about models, benchmarks, and infrastructure, but not nearly enough time talking about the people building them and the communities around them.\n\nThe discussion, led by my WiMLDS colleague Marie Sacksick and Joanna Kramer, Co-Founder at WISE (Women in Safety & Ethics), revolved around two main themes:\n\nWhat I appreciated most is that the conversation stayed practical. Rather than stopping at broad principles, it encouraged everyone to think of one concrete action they could take within their own community to move things forward.\n\nSeeing this kind of discussion alongside deep technical workshops felt important. Open source is ultimately about people as much as it is about code.\n\nAnd that’s a wrap.\n\nTo sum it up, GOSIM 2026 was one of my favourite conferences this year. It’s highly technical, low on commercial fluff, and driven by a genuine commitment to advancing open source. I loved the diversity of topics, the deeply technical themes, and the formats, with workshops long enough to actually go somewhere.\n\nLooking back, a few things stood out. Infrastructure is becoming a real pillar of the AI stack, agentic AI** **is clearly everywhere, and it was refreshing to see data science still very much part of the picture alongside all the LLM hype.\n\nThank you for reading, and feel free to share your thoughts in the comments! 😊\n\n[GOSIM Paris: This Is What Open Source AI Looks Like in 2026](https://pub.towardsai.net/gosim-paris-this-is-what-open-source-ai-looks-like-in-2026-91b6d0bd702a) was originally published in [Towards AI](https://pub.towardsai.net) on Medium, where people are continuing the conversation by highlighting and responding to this story.", "url": "https://wpnews.pro/news/gosim-paris-this-is-what-open-source-ai-looks-like-in-2026", "canonical_source": "https://pub.towardsai.net/gosim-paris-this-is-what-open-source-ai-looks-like-in-2026-91b6d0bd702a?source=rss----98111c9905da---4", "published_at": "2026-06-26 20:01:01+00:00", "updated_at": "2026-06-26 20:40:45.562759+00:00", "lang": "en", "topics": ["large-language-models", "ai-research", "ai-tools", "ai-infrastructure"], "entities": ["GOSIM", "Station F", "Sir Gowers", "Gaël Varoquaux", "Probabl", "scikit-learn", "vLLM", "Red Hat"], "alternates": {"html": "https://wpnews.pro/news/gosim-paris-this-is-what-open-source-ai-looks-like-in-2026", "markdown": "https://wpnews.pro/news/gosim-paris-this-is-what-open-source-ai-looks-like-in-2026.md", "text": "https://wpnews.pro/news/gosim-paris-this-is-what-open-source-ai-looks-like-in-2026.txt", "jsonld": "https://wpnews.pro/news/gosim-paris-this-is-what-open-source-ai-looks-like-in-2026.jsonld"}}