Changes coming to H2O-3 open source

Where H2O-3 stands today

3.46.0.11 is the current stable release and H2O-3 is now covered under our FedRAMP High certification.

This has streamlined on-premise and air-gapped deployments for various agencies. Furthermore, our vulnerability patching workflow complies with FedRAMP standards, ensuring documented intake, triage, remediation, and disclosure within designated timeframes.

What's landing in H2O-3 3.46.0.12

We are sharing details in advance regarding several critical changes coming to H2O-3. We believe you should be fully apprised of these updates before upgrading, rather than discovering them after the fact.

Two real changes are landing in 3.46.0.12.

Four components are being removed from H2O-3 Open Source and will be leaving:

• The MOJO runtime
• Hadoop
• Kubernetes deployment
• SparklingWater (Spark)

Beginning with the 3.46.0.12 release, these four components are scheduled to be removed from the open-source distribution channels and will move to H2O-3 Secure, where they will remain fully supported and accessible. These features are mainly beneficial for large enterprise environments and are less vital for the general community, though we welcome your feedback to the contrary in the GitHub discussion.

We've added telemetry.

Telemetry is opt-in and off by default and set to DO_NOT_TRACK=1, more clarity on how we structured the ability to disable can be found at https://donottrack.sh/. The complete list of what is and isn't collected is in the repo README. For clarity, we don't track your code, your data, column names, hostnames, or paths. We'll proactively announce and update the H2O-3 GitHub README before making any change to what's collected.

Why we're making these changes

These four components are essential for running H2O-3 as production infrastructure at scale for commercial applications. However, they introduce significant engineering commitments that draw attention away from our OSS Machine Learning and Feature Engineering goals. Thus, they are better suited for a dedicated production tier.

Vulnerability response has become a specialty in itself. Around 40,000 CVEs were published in 2024, roughly 46,000 in 2025, and the first half of 2026 is running about 50% ahead of the first half of last year. 

NIST restructured how the NVD operates in April to keep pace. Better tooling is finding real bugs, which is good. The problem is the gap it opens between finding and fixing. 

JP Morgan's Patchmageddon put numbers on it this month, citing Anthropic data: of 530 high and critical vulnerabilities reported to maintainers, only 75 had been patched. Meanwhile the average time between disclosure and first exploitation has fallen to zero days, and in around 60% of breaches a patch already existed when the compromise happened.

The majority of CVEs affecting a JVM data platform do not originate within our own codebase; instead, they reside in the underlying dependency trees. Integrations with Hadoop and Spark pull in extensive transitive dependencies, meaning every security advisory across these trees comes with a strict timeline attached; driven by FedRAMP windows, SOC 2 commitments, and contractual customer terms.

By consolidating commercial and production-level components into H2O-3 Secure, we can deliver the rigorous vulnerability management, consistent patching schedules, and long-term maintenance that these features require. This shift also enables us to keep the open-source distribution centered on what our community utilizes most heavily: AutoML, the comprehensive algorithm library, single-node experimentation, and Python and R clients.

Over the past two years, the same two themes have come back through support tickets, GitHub issues, and user calls. Running H2O-3 in commercial production: 

• Scale and reliability for production workloads 
• Enterprise class authentication and role-based access control 
• Audit logging and reporting 
• Workload isolation and identity management 
• Deployment orchestration and dynamic scalability
• Integration with production data sources
• Documented vulnerability management 
• Incident response timelines 
• Vendor risk documentation 
• Audit evidence for SOC 2, ISO 27001, and ISO 42001

To address these demands, we introduced H2O-3 Secure:

• Commercially supported and built with the rigor of H2O.ai's Cloud for Government (H2O-GOV)
• Brings a FedRAMP-grade security posture to production workloads
• Features multi-node deployment, prioritized CVE patching, and audit support
• Includes the MOJO runtime and native integrations for Databricks, Kubernetes, Hadoop, and Spark

What stays open source

• Apache 2.0 will continue to serve as the license for open source H2O-3, which applies to all upcoming releases we provide.
• All previously published releases remain accessible. Checksums, licenses, and URLs will stay exactly the same. Version 3.46.0.11 is not going anywhere, pinning your builds to it will ensure everything keeps running smoothly.
• Core capabilities of the H2O-3 platform: Industry leading AutoML, the complete optimized math and algorithm libraries, Python and R clients, the REST API, and community support for single-node and local clusters.
• For 3.46.0.11 and prior releases: the Hadoop and build docs remain in the README with notes about which versions they apply to.
• Regarding SparklingWater: Locked to version 3.46.0.6 and all prior releases remain available along with the build documentation and associated README files.

"Apache 2.0 forever" doesn't mean much without saying what maintenance comes with it, so:

• Critical-severity fixes, and anything under active exploitation, ship to open source immediately once we have a fix.
• All other security fixes land in open source on our standard patch release cycle.
• New features and capabilities will be released on a delayed schedule.

What we're asking

We want to make this transition smooth for the community. If you're affected, we'd like to hear from you:

1. If you're running MOJO in production, what's it scoring, and on what? I'm looking for a clearer picture of real-world usage and deployments.
2. Is there something in this blog that honestly doesn't work as a landing place for you? I would like to hear about that as well.
3. If you teach with H2O-3, or your paper's reproducibility section points at this repo, then let us know. We will support your use case, please reach out!
4. Are you using H2O-3 via Aquarium? How can we improve and make the UX better, let us know!
5. If you need H2O-3 Secure functionality and are non-profit, please reach out!

Reply in GitHub Discussions. We'll be paying attention. If you'd rather not post publicly, send us an email to OSS@h2o.ai.

If you're an individual data scientist, student, academic, or non-profit needing any of the four removed components, send an email to us with details on what you're building. We'll be sure to respond and work to support you.

For security related concerns please reach out to enterprise@h2o.ai or support@h2o.ai.  More information available on our GitHub repo.

To the community

While the committers list in our README contains around 100 names, it significantly underrepresents our actual community. H2O-3 owes its success to those who reported bugs at 2 AM, sent unsolicited patches, utilized it in classrooms, and challenged us in the issues.

Democratizing AI and AI for Good wasn't a slogan we added later. It's why H2O-3 was open source from the start, and it's why the algorithm library, AutoML, and single-node development capabilities remain fully open source going forward.

Please reach out with any questions.