Flower

At Flower AI Summit 2026, Flower Labs' own Charles Beauville introduced Flower Agents and Project Kaya, and made a case for what separates a good agent from a great one. His framework: a good agent needs context, access, and control. Context so it understands the task. Access so it can act on it. Control so it can actually be trusted in production. But a great agent needs one more thing: collaboration. Isolated agents break down the moment a problem crosses team or organizational boundaries. Flower Agents are built to work together across those boundaries, without breaking privilege isolation, and without any new infrastructure. SuperGrid already provides the orchestration, communication, isolation, and auditability needed to run collaborative agents by design. Project Kaya has already been deployed internally at @flwrlabs, triggering @github fixes from Slack, creating Notion docs from live context, and running federated analytics in natural language across a federation of organizations. The full talk is coming to @YouTube soon. Join the waitlist. See link in thread.

FlowerTune is our federated fine-tuning framework. The leaderboard that comes with it lets you benchmark LLMs across 4 domains where data privacy is key: General NLP, Finance, Medical, and Code. Getting started is straightforward. Pick a challenge, implement your fine-tuning approach on top of the provided template (model-agnostic, works with any base model up to 14B parameters), run evaluation, and submit. The framework handles federated simulation out of the box via flwr run, so you don't need a distributed setup to get going. The benchmark has already attracted contributions from teams at @ZJU_China, @penn_state, @AlibabaGroup, Gachon University, @SonyAI_global, and many others, and was presented at NeurIPS 2025. Core developer is @yangao381. If you're working on LLM fine-tuning and want to understand how your approach holds up in realistic federated conditions, this is the place to test it: buff.ly/y9fA9zg

At Flower AI Summit 2026, @dstripelis_ presented how to work with us via the selective Flower Pilot Program, as well as past Pilot success stories. From a federated network spanning 108 hospitals for early infectious disease detection, to cross-bank credit risk modeling where no customer data ever leaves its source. These are real projects with concrete outcomes! Together, we connect siloed data, compute, and models, and jointly ensure that these systems are production ready. That's what the Flower Pilot Program is built for. Now in Batch 4, the program guides organizations through 3 phases: (1) problem scoping, (2) integration, and (3) live deployment. This is all done in an accelerated 12-week sprint, working directly alongside the core Flower Labs team. Past exits include @SYNTHEMA_EU (a multi-hospital European consortium), AI4Cosmetics, University of Maryland, and others across healthcare, finance, and beyond. If your organization is ready to move from exploration to execution on federated AI, applications for Batch 4 are open: buff.ly/Oy8ijCD

One of the unique capabilities of Flower Hub is FAB (Flower App Bundle). First detailed in our blog on 21 May 2026, FAB allows any federated system to migrate to new versions of Flower without requiring any code changes. Until now, evolving the platform risked breaking apps already published and used. The stable FAB format solves that with a single field: fab-format-version. It tells Flower which app format and validation rules to apply during build, including metadata and FAB content validation. For Flower Hub to function as a shared ecosystem (not just a file host), publishers need a common contract. FAB is that contract: consistent validation, predictable packaging, and a clear path forward as the platform grows. The practical result is a clearer boundary for publishers: backward compatibility for older apps, while newer apps can adopt richer metadata and stricter validation when they are ready. If you publish Flower Apps, this is the field to review before you republish on Flower Hub. Read the blog post (written by Mohammad Naseri and @yangao381) and reply here with questions: buff.ly/7ysOVlC

Flower 1.30 is out. The main highlight in 1.30 is the new task system for executor process lifecycle management. Tasks are now tracked objects with explicit lifecycle states (e.g. created, active, completed), giving the framework real visibility into what's running and a reliable foundation for orchestration, for both flwr-serverapp and flwr-clientapp. Run status is now tied directly to its primary task, giving the orchestration layer a cleaner and more reliable foundation. Two additional 1.30 features include: (1) runtime version compatibility checks now enforce matching major.minor Flower releases across AppIo communication paths, preventing incompatible combinations from proceeding silently; and, (2) TLS support has been added for the ServerAppIo and ClientAppIo APIs, with root certificate configuration available for flwr- commands and flower-superexec. Thanks to the contributors who made this release possible. Full announcement here, written by Heng Pan: buff.ly/srToCQe

Tony James, Chief Architect at @RedHat, joined us at Flower AI Summit 2026 to talk about something very important: what it actually takes to run federated AI in enterprise environments. His talk focused on regulated, air-gapped, multi-network, heterogeneous-hardware environments where trust boundaries are real constraints and opening a firewall rule is a security conversation. He walked through 3 architecture patterns: using Flower as the coordination layer across execution environments; handling heterogeneous hardware footprints (HPC, cloud, edge, @AMD/@intel) with @kubernetesio as the unifying platform; and moving connectivity up to the application layer rather than bridging networks at the infrastructure level. This, all using @RedHat Service Interconnect to establish point-to-point links between Flower's SuperLink and SuperNode without broadening the security boundary. The concrete output of the @RedHat x @flwrlabs collaboration is a Flower container image built on @RedHat Universal Base Image, so enterprises running RHEL in FIPS mode can inherit those attestations through to the federated workload. Full talk on YouTube now: buff.ly/frBhXPg

Nisal Hemadasa from Hamburg University of Technology came to Flower AI Summit 2026 to talk about something very important: concept drift. One type of existing approach handle it by clustering clients into groups and assigning each a separate global model. It works, until the number of clients and drift types keeps growing. Then the model count explodes and so does the cost. Nisal's approach is different. Rather than clustering, it identifies which layers of a neural network are most stimulated by a given type of drift, shares only the less-stimulated layers globally, and keeps the rest personalized per client. This results in one global model, better scalability, and more robust drift recovery. The full talk is now on @YouTube: buff.ly/bsYyRdL

From launch day to 124 apps in under two months. Flower Hub keeps growing as 5 new apps added in the last few days alone. Week after week, the community is contributing new projects, and the ecosystem is becoming more practical, more reproducible, and more useful for real-world federated AI development. To everyone who has published an app, opened a PR, or shared feedback: thank you. This is your work. If you are building in federated, decentralized, or privacy-preserving AI, Flower Hub is worth a close look. Check it out here: buff.ly/H13dUqv

Federated learning keeps data local. But what about the model updates themselves? Mohammad's latest app on Flower Hub tackles exactly that, combining federated learning with CKKS homomorphic encryption via TenSEAL (built on @Microsoft SEAL), so the server aggregates encrypted updates and decrypts only the final result. Individual client contributions stay private end to end. Getting started takes 2 commands: flwr new (at)mohammad/he-flower-example flwr run . --run-config "mode='he_ckks'" This is what Flower Hub is for; making advanced privacy techniques concrete and runnable. Check it out here: buff.ly/6fWVsMP

Federated AI needs standards, which is why @flwrlabs is joining @MLCommons to help build them. Together, we're bringing federated AI into the infrastructure conversation that determines how collaborative and agentic AI actually gets deployed at scale, especially in the domains where data privacy is non-negotiable. The first concrete result: MedPerf is officially integrating with Flower. Researchers can run federated clinical AI studies across institutions without raw patient data ever moving. As agentic AI systems move into high-stakes domains like healthcare, the ability to benchmark and train across institutions without centralizing sensitive data becomes foundational. The MedPerf project, first described in a 2021 paper and published in @Nature Machine Intelligence in 2023, was built precisely for this: bringing models to the data, not the other way around. Flower makes that workflow accessible at scale. Our investment in federated healthcare goes back years. This integration is the next step and if you’re interested, you can read more about it here: buff.ly/xGmjXvl If you work in federated AI, healthcare AI, or large-scale distributed systems, you should definitely take a look at above integration, as well as head over to flower.ai to read more about what we’re building at @flwrlabs. If you’re interested in news like this, don’t forget to follow us as it is the best way to stay up-to-date with the latest developments in federated AI.

Tabular data exists everywhere and there's a federated learning app for it on Flower Hub. @chongshenng/fed-tab (created by @chongshenng -- Research Scientist at @flwrlabs) demonstrates federated classification on the Forest Covertypes dataset using TabPFN-2, @prior_labs' transformer-based foundation model for structured data. The dataset consists of 30×30m patches of US forest land, each labeled with the dominant tree species; the kind of prediction task that shows up constantly across industry. In the app, data is partitioned across clients and trained over several rounds, without any raw data ever leaving each node. The timing is notable as @prior_labs (the team behind TabPFN) just announced a definitive agreement to be acquired by @SAP, with over €1B in intended investment over 4 years. TabPFN-2.6 currently sits atop the TabArena benchmark for tabular foundation models. What fed-tab shows is what becomes possible when a model like that runs inside a federated framework: strong tabular predictions, without centralizing the underlying data. You can try it yourself in three commands: pip install flwr flwr new @chongshenng/fed-tab flwr run . and you can find more about it on buff.ly/LZEPcZk

@daniel_janes The full keynote can be found here on @YouTube: buff.ly/9zixgnn Let us know in thread what you're most excited about when it comes to federated/decentralized AI, and don't forget to follow us for news and announcements.

@daniel_janes 3. Project CAIA is a collaborative AI agent that operates across organizational boundaries via SuperGrid. A sneak peek at what agentic AI looks like when it can actually collaborate.

The data that can't be centralized (sits inside organizations, governed, sensitive, fragmented) is 133x larger. The next era is here. The opportunity is evident: data used to train today's best models represents just 15 trillion tokens. [thread]

Flower recently crossed 2M downloads. After reaching 1M downloads recently, the next 1M came in just 7 months. We're grateful for everyone choosing to work and build with Flower, and look forward to the road to 3M downloads. Collaborative, distributed, and privacy-preserving AI is no longer a niche area of interest. More teams are now approaching it as a practical foundation for how AI systems should be built, deployed, and governed. We are seeing this across the Flower ecosystem: in research, enterprise deployments, open-source projects, and production-oriented tooling for teams that need AI to operate across distributed environments while maintaining control over data, infrastructure, and governance. This milestone is a big shoutout to the broader Flower community! Thank you to everyone contributing, deploying, experimenting, and advancing this field with us. If you like this sort of content, don't forget to follow us here.

Recording can be found here on @YouTube: buff.ly/pGsXSeo

A key finding worth noting: the same device behaves differently depending on the task. Image classification and LLM fine-tuning place different demands on hardware, meaning trace-based scheduling can produce quietly misleading evaluations. BouquetFL makes those differences visible. BouquetFL is available on the Flower Hub: buff.ly/442FE0M

At Flower AI Summit 2026, Arno Geimer (PhD Student, University of Luxembourg) addressed a challenge in federated learning research: how do you evaluate client heterogeneity at scale without access to dozens of physical devices?