Skip to main content
Version: v1.7.0

Release notes

v1.7.0 | August 19, 2026​

Overview​

H2O Hydrogen Torch v1.7.0 simplifies the model deployment workflow and improves H2O MLOps scoring output. This release replaces in-app deployment management with a new upload-to-MLOps workflow, adds a new output format for MLOps scoring pipelines, and includes several UI and UX improvements. The major points of this release are as follows:

  • Simplified upload-to-MLOps workflow: In-app deployment management has been replaced with a new model upload workflow; deployments are now created and managed entirely in the H2O MLOps UI
  • Native MLOps scoring output: A new opt-in setting returns predictions from the MLOps scoring endpoint as native JSON objects instead of JSON-encoded strings
  • Always-visible list checkboxes: Row-selection checkboxes are now always visible on all list pages, replacing the previous two-step selection flow
  • Full-width layout: The UI now uses the full browser viewport width
  • Updated GPU stack: CUDA 13.0 with torch 2.13.0, upgraded from CUDA 11.8 with torch 2.7.1; now requires NVIDIA driver version 580 or later

Experiments​

  • New: H2O Hydrogen Torch now lets you upload a trained model to an H2O MLOps Workspace from the experiment's MLOps tab, previously named Deployments. To learn more, see Upload model to H2O MLOps.
    • Why?: Deployments are now created and managed in H2O MLOps, where you can also configure scoring runtimes, security, monitoring, and Kubernetes resources.
  • New: A new Native MLOps output training setting, available when Build scoring pipelines is turned on, returns each prediction as a native JSON object instead of a JSON-encoded string. To learn more, see Native MLOps output.
    • Why?: Native output removes the json.loads() step that a scoring client would otherwise need. The setting is off by default and requires H2O MLOps 1.2.0 or later.

UI & UX​

  • Improvement: Row-selection checkboxes are now always visible on all list pages (datasets, experiments, predictions, and projects). The delete, stop, and compare actions apply directly to the selected rows.
    • Why?: The previous two-step flow required you to click an action button and then select rows in a separate view.
  • Improvement: The UI now uses the full browser viewport width.
    • Why?: The previous fixed-width layout left unused space on wider monitors.
  • Improvement: UI setting labels now spell out acronyms and use sentence case. For example, Nms Iou Threshold is now Non-maximum suppression (NMS) IoU threshold.
    • Why?: Spelling out acronyms and using sentence case make setting labels easier to read.

Infrastructure​

  • Improvement: H2O Hydrogen Torch now runs on CUDA 13.0 with torch 2.13.0, upgraded from CUDA 11.8 with torch 2.7.1.
    • Why?: Later CUDA and torch releases add support for current GPU hardware. H2O Hydrogen Torch and the downloadable Python scoring pipeline now require NVIDIA driver version 580 or later.
  • Improvement: H2O Hydrogen Torch now uses transformers 5.12.1, upgraded from 4.57.6.
    • Why?: The upgrade gives you access to additional model architectures. If you use a custom Hugging Face backbone, confirm that it still loads with transformers 5.x.

Deprecations​

  • Deprecation: In-app deployment management has been removed, including the deployment list, the deployment logs, and the Allow custom MLOps limits setting.
    • Why?: Deployments are created and managed in H2O MLOps. Deployments that you created with an earlier version keep running there, but H2O Hydrogen Torch no longer displays them: in-app deployment records are not carried over to the new upload records.
  • Deprecation: The Deployed experiment status has been replaced with the Uploaded status.
    • Why?: H2O Hydrogen Torch now tracks the models you uploaded, not the state of their deployments. The new status is derived from upload records, so an experiment that previously showed Deployed shows Finished until you upload it again.
  • Deprecation: The JSON-encoded string output of the H2O MLOps scoring pipeline is deprecated. Turn on the Native MLOps output training setting to move to the new format.
    • Why?: Native output becomes the default in v1.8.0, and the JSON-encoded string format is removed in v1.9.0. Update any scoring client that calls json.loads() on the output field before you upgrade to v1.8.0.

Bug Fixes​

  • Fix: Fixed a crash in EfficientDet object detection when a model returns fewer detections than Maximum detections per image allows.
  • Fix: Fixed the Could not read image error for images that OpenCV cannot decode.
  • Fix: Fixed a deadlock in the mean average precision (mAP) calculation for object detection that could halt training.
  • Fix: Fixed the table download button appearing for a table that contains no rows.
  • Fix: Fixed the downloadable Python scoring pipeline failing to build because its Dockerfile installed Python 3.10.
  • Fix: Fixed text span prediction not scoring the last rows of the input in the downloadable scoring pipeline and in deployed models.
  • Fix: Fixed deployed audio models scoring a truncated signal when Inference chunk method is set to Varying.

Documentation​

  • New: All new features and settings for v1.7.0 have been documented.

v1.6.0 | January 31, 2026​

Overview​

H2O Hydrogen Torch v1.6.0 focuses on infrastructure modernization, security compliance, and code base maintenance. This release brings FedRAMP compliance through security hardening, migrates to Python 3.12, and streamlines the platform by removing deprecated features. The major points of this release are as follows:

  • FedRAMP compliance: Chainguard base image and CVE mitigation for enhanced security
  • Infrastructure improvements: Demo datasets are now served via CDN for improved performance and reliability
  • Python 3.12: Upgraded code base from Python 3.10 to Python 3.12
  • Deprecations: Removed Neptune logging (deprecated service) and Graph modeling problem types

To learn more about the new release, observe the below subsections.

Infrastructure​

  • Improvement: Demo datasets now use CDN instead of the AWS S3 bucket directly.
    • Why?: This improves download performance and reliability when accessing demo datasets.
  • Improvement: H2O Hydrogen Torch now uses a Chainguard base image with CVE mitigation for FedRAMP compliance.
    • Why?: This ensures the platform meets federal security requirements and reduces vulnerability exposure.
  • Improvement: The H2O Hydrogen Torch code base has been upgraded from Python 3.10 to Python 3.12.
    • Why?: This brings performance improvements and access to the latest Python features and security updates.

Deprecations​

  • Deprecation: Neptune logging has been removed.
    • Why?: The Neptune service has been deprecated. Users should use alternative logging integrations.
  • Deprecation: Graph modeling problem types have been removed, including:
    • Graph node classification
    • Graph node regression

Bug Fixes​

  • Fix: Resolved an issue with label handling in CustomDataset.
  • Fix: Fixed a data import crash that occurred when label columns contained non-string values.

Documentation​

  • New: All new features and settings for v1.6.0 have been documented.

v1.5.0 | January 27, 2024​

Overview​

At last, the release of H2O Hydrogen Torch v1.5.0 brings significant enhancements. It introduces support for a new problem type, furthering our mission to provide a no-code, accessible platform to train state-of-the-art deep neural networks (models) across diverse problem types. Whether you're a seasoned data scientist or have no coding experience, H2O Hydrogen Torch v1.5.0 makes building cutting-edge models easier and more efficient. The major points of this release are as follows:

  • New problem type: Multi-modal causal language modeling

    • H2O Hydrogen Torch v1.5.0 now supports multi-modal causal language modeling, enabling the creation of models around this problem type. These models generate textual responses based on both textual and visual inputs, such as queries paired with images.
    • Why? This feature empowers users to tackle complex tasks requiring cross-modal data understanding.

  • Automatic deep learning (AutoDL)

    • H2O Hydrogen Torch v1.5.0 introduces AutoDL. This new feature allows you to train a set of models with just one input: a time budget, which specifies the number of experiments that H2O Hydrogen Torch should generate, each with different values for specific hyperparameters referred to as grid search hyperparameters.
    • Why? This feature automates the model-building process, making it especially useful for non-data scientists who need to create high-performing models quickly and efficiently. Furthermore, this feature lowers the barrier to entry for deep learning and enables anyone to build "good enough" models without extensive expertise.

  • Quantization Dtype (data type)

    • A new setting for text-only problem types is available during QLoRA training named Quantization Dtype.
    • Why? This new setting allows users to control the level of quantization applied to the model, balancing memory efficiency and precision.
Feedback