Skip to main content
Version: v0.23.x

Release notes

Version 0.23.0 (October 1st, 2026)​

New Features​

  • [Training] Text classification now supports BERT-family encoder backbones in addition to decoder backbones. BERT, RoBERTa, XLM-RoBERTa, and ModernBERT models can be deployed through Model Zoo. DeBERTa and DistilBERT models can be trained and used for predictions, but not deployed yet.
  • [Experiments] When a model hub is configured, the backbone dropdowns on the new experiment and AutoML pages now list the models the hub holds. H2O Cascade's default models that are in the hub appear under Pretrained models (Available Backbones on the AutoML page). All other models in the hub appear under Other models in model hub. If the hub can't be reached or serves none of the defaults, a notice explains why.
  • [Assistant] When an administrator selects the Assistant Model on the Settings page, H2O Cascade now checks whether the model's endpoint supports tool calling and shows the result on the Assistant Model setting. If the endpoint doesn't support tool calling, the assistant now returns an error instead of answering without its tools.
  • [Platform] Added an option to have each training worker exit after it finishes one job (H2O_LLM_STUDIO_CELERY_ONE_SHOT), for autoscaled deployments that start one worker per training job.

Fixes​

  • [Model Zoo] Fixed deployments getting stuck in "starting", with a GPU still reserved, if they were created or restarted while Model Zoo was idle and you left the page before they started.
  • [Training] Fixed classification training failing when the validation split didn't contain every class.
  • [Training] Fixed classification experiments with integer labels (for example, 0, 1, 2) failing with an unclear error when validation or prediction data contained a label not seen during training. The error now lists the unknown labels.
  • [Experiments] Fixed object detection boxes in validation results being drawn at the wrong scale for images larger than 512 pixels.
  • [Platform] Fixed pages reloading on their own or stopping you from moving to another page, and fixed signing in again taking you to the home page instead of the page you were on.

Security​

  • Updated the PyJWT library to fix several vulnerabilities.

Version 0.22.9 (September 22nd, 2026)​

Fixes​

  • [Data Generation] Fixed Check Config and Start running with no model selected and then showing an unclear error. Both buttons are now disabled until a model is selected, and Reset Settings picks a model from the allowed list.

Security​

  • Updated Model Zoo's inference runtime (vLLM) to fix a vulnerability that let a crafted request use up GPU memory and crash the deployment.
  • Updated libraries across H2O Cascade's services to fix several vulnerabilities, including one that allowed TLS certificate spoofing.

Version 0.22.8 (September 18th, 2026)​

New Features​

  • [Platform] The Keycloak admin role name is now configurable (KEYCLOAK_ADMIN_ROLE, default llmstudio_admin), so you can use an existing role in your identity provider instead of creating a new one.

Fixes​

  • [Assistant] Fixed assistant attachments larger than 10 MB failing to upload, even though the UI allows up to 20 MB.
  • [Assistant] Fixed the assistant rejecting models whose names start with openai/ (for example, openai/gpt-oss-120b) when the endpoint lists them that way.
  • [Data Generation] Fixed data generation still offering models the configured endpoint doesn't serve after an administrator clicked Refresh Models, which caused those runs to fail with "Invalid model".
  • [Platform] Fixed System Status showing Object Storage as unhealthy on external S3 installs where the S3 role can only access H2O Cascade's own buckets.
  • [Platform] The About page and the assistant now use the H2O Cascade name.

Version 0.22.7 (September 14th, 2026)​

Fixes​

  • [Platform] Fixed a single failed request making the UI retry it endlessly and show the same error message over and over.
  • [Platform] Fixed experiment and data generation configurations failing to load with "Configuration file not found" on external AWS S3 storage once the stored download link expired.

Security​

  • Updated web UI dependencies to fix six security advisories.

Version 0.22.6 (September 11th, 2026)​

Fixes​

  • [Model Zoo] Fixed deployments failing to start on NVIDIA Hopper and Blackwell GPUs (for example, H100 and B200).
  • [Model Zoo] Fixed deployments failing with a permission error on OpenShift, which runs containers under a random user ID.
  • [Model Zoo] Fixed Model Zoo crashing at startup on FIPS-enabled hosts.
  • [Model Zoo] Fixed the Llama Guard model in Model Zoo returning an error on every request.
  • [Experiments] Fixed experiments staying "queued" forever when the training job failed, for example, when the training worker stopped unexpectedly or hit a time limit. They are now marked as failed, with the error in the logs.
  • [Datasets] Fixed adding a URL-based demo dataset hanging and then showing a generic error when the download host isn't reachable. It now fails quickly with a message that names the host.
  • [Data Generation] Fixed data generation jobs with an empty User Prompt Template reporting success while every row contained the same generic greeting. Start and Check Config are now disabled until a template is entered.

Security​

  • Updated the Transformers library to fix a path traversal vulnerability.
  • Updated the web UI framework (Next.js) to fix two security advisories.
  • Removed an unused copy of pip from the training and Model Zoo images. Security scanners were reporting setuptools vulnerabilities from it.

Version 0.22.5 (September 2nd, 2026)​

New Features​

  • [Experiments] Added an opt-in option to compute feature importance for TabH2O tabular predictions, shown as a ranked chart in the experiment's results.
  • [Model Zoo] Added a provisional /api/v1 endpoint namespace for agent-facing API clients, with an OpenAPI 3.1 spec and a capability manifest. It currently covers the existing Model Zoo endpoints.

Fixes​

  • [Platform] Fixed the assistant, data generation, and AutoML model dropdowns showing no available models when configured against an OpenAI-compatible gateway instead of OpenAI itself. Every enabled model is now offered to every feature, instead of only the overlap with each feature's built-in list.

Security​

  • Fixed a vulnerability in how the assistant's web-fetch tool parses robots.txt files, which could let a malicious site hang the request with a crafted file.

Version 0.22.4 (August 31st, 2026)​

New Features​

  • [Platform] Added support for the OpenAI gpt-5.6 model family, available in the assistant, data generation, and AutoML.
  • [Training] Image classification training now supports non-square (rectangular) image resolutions and an explicit resize mode (squash, pad, or crop), instead of always inheriting the backbone's default center-crop behavior.

Fixes​

  • [Assistant] Fixed document parsing (PDF, DOCX, PPTX, HTML) failing in airgapped environments by bundling the required OCR and layout model weights instead of downloading them on first use.

Security​

  • Updated Model Zoo's inference runtime (vLLM and PyTorch) to fix a memory-corruption vulnerability.

Version 0.22.3 (August 18th, 2026)​

Fixes​

  • [Training] Fixed training.image_resolution being ignored for timm-based vision backbones in image classification, so models now train at the resolution you configure instead of silently falling back to the backbone's default.
  • [Training] Fixed image classification and object detection models training with reduced numerical precision and skipping weight decay on normalization layers, both of which were quietly limiting final model accuracy.

Security​

  • Requests made with the internal service token but no identified user are now rejected instead of running with full administrator access.

Version 0.22.2 (August 14th, 2026)​

New Features​

  • [Platform] Dataset uploads to Azure Blob Storage now upload directly from the browser, matching the existing direct-upload behavior for S3-compatible storage.
  • [Platform] The task-acknowledgment option added in 0.22.1 (H2O_LLM_STUDIO_CELERY_TASK_ACKS_LATE) now also applies to the AutoML and data generation workers, in addition to the training worker.

Security​

  • Disabled tracing in the assistant's underlying agent framework, which by default sent prompts and responses to a third-party endpoint.

Version 0.22.1 (August 11th, 2026)​

New Features​

  • [Platform] Added an option to change when the training worker acknowledges queued jobs (H2O_LLM_STUDIO_CELERY_TASK_ACKS_LATE), so autoscalers can see training jobs that are still in progress.

Version 0.22.0 (August 7th, 2026) — H2O Cascade​

Rebrand​

H2O Enterprise LLM Studio is now H2O Cascade. The product has grown beyond fine-tuning LLMs to support additional modalities and use cases — including image classification and object detection — with more on the way. This rebrand reflects that broader scope; the platform itself, your existing projects, and your data are unaffected.

Documentation for previous H2O Enterprise LLM Studio releases remains available via the version switcher.

New Features​

  • [Platform] Added native Azure Blob Storage support as an object storage backend.
  • [Experiments] New experiments can now start directly from a previously trained experiment's weights (for example, continuing SFT training into RL) instead of a manual download and re-upload of the model.
  • [Training] RL experiments now report a validation score at step 0, in addition to during training.

Fixes​

  • [Assistant] Fixed the assistant failing when chatting against OpenAI-compatible endpoints, and improved the error shown when a response fails to stream.
  • [Training] Fixed a regression that slowed down object detection training.
  • [Training] Fixed RL experiments getting stuck showing "validating" through the first half of a run.
  • [Training] Fixed training runs being marked as failed — and their logs discarded — due to a transient error while monitoring worker processes.
  • [Experiments] Fixed the training-curve chart getting flattened by a single loss outlier spike.
  • [Experiments] Fixed the "System Prompt Column" selection on the new experiment page being dropped from the submitted experiment configuration.
  • [Data Generation] Fixed data generation jobs hanging in airgapped environments while trying to download tokenizer files.
  • [Model Zoo] Fixed deployments failing to wake up when GPU capacity was briefly unavailable; they now wait and retry instead.
  • [Platform] Fixed broken dataset uploads to external object stores caused by incorrect CSP and presigned URL settings.

Security​

  • Fixed an authorization gap that could allow deploying an experiment from a project the user does not have access to.

Version 0.21.2 (July 29th, 2026)​

New Features​

  • [AutoML] AutoML logs now show live experiment progress — status transitions, periodic runtime/ETA/loss updates, and a summary when each experiment finishes or fails.

Fixes​

  • [AutoML] Fixed AutoML showing reinforcement learning (RL) experiments as regular causal LM experiments, including an irrelevant output column option for them.
  • [AutoML] Fixed AutoML runs getting stuck showing "running" when a failure occurred partway through.
  • [AutoML] Fixed LLM-as-a-Judge being rejected as the optimization metric when starting AutoML through the assistant.
  • [AutoML] RL experiments in AutoML now always use LLM-as-a-Judge as the optimization metric, since no other metric applies. The metric selector is hidden for RL.
  • [Training] RL rollout prompts and completions are no longer written to training logs by default, preventing training data from leaking into logs.
  • [Model Zoo] Fixed an issue that prevented deployment of RL models.
  • [Datasets] Fixed CSV row counts being overstated for datasets containing multi-line fields.
  • [Datasets] Removed an unused output column from the RL demo dataset that incorrectly implied labels are required for RL training.
  • [Assistant] Fixed the assistant failing when using model names from a gateway's model catalog.

Version 0.21.1 (July 27th, 2026)​

New Features​

  • [Platform] Support for running against a customer-managed external Postgres database (for example, AWS Aurora), including optional IAM authentication.

Fixes​

  • [Platform] Improved performance of the experiments, deployments, and data generation list pages, which no longer load full task logs on every automatic refresh.
  • [Data Generation] Fixed an issue where data generation jobs could get stuck showing "running" instead of completing or failing.
  • [Training] Fixed a crash when training on datasets containing missing values.
  • [Experiments] Fixed a server error that broke the experiment page and blocked AutoML from starting the next experiment for experiments trained on JSON datasets.
  • [Platform] Fixed the settings page so a failed OpenAI models fetch no longer blocks the rest of the page from loading.

Version 0.21.0 (July 20th, 2026)​

New Features​

  • [Model Zoo] Model deployments now run through a shared, GPU-aware deployment manager instead of a dedicated worker per deployment, improving scalability and cost efficiency.

Version 0.20.2 (July 9th, 2026)​

New Features​

  • [Model Zoo] New batch inference API — submit batch jobs asynchronously, poll or cancel them, and stream progress or receive a webhook callback on completion. Batch jobs are priority-queued so they can't starve interactive traffic, with native batched support for vLLM, chat, NER, PII masking, the Danube3 classifier, and BART.
  • [Platform] Storage bucket names are now configurable via environment variables.

Fixes​

  • [UI] Fixed a dropdown selection issue affecting components using multi-select.
  • [Training] Fixed a training failure that occurred when input CSV files used CRLF line endings.
  • [Platform] Resolved a dependency conflict that could block installation.

Security​

  • Fixed server-side request forgery (SSRF) findings in frontend server actions and API routes.

Version 0.20.1 (July 2nd, 2026)​

New Features​

  • [Data Generation] You can now choose the output dataset's problem type when generating data.
  • [Platform] Further upgraded the core ML stack: PyTorch 2.11, vLLM 0.22.1, and CUDA 13 / cuDNN 9 base images. This raises the minimum required NVIDIA driver version.
  • [Model Zoo] Added per-model request queueing with backpressure, along with cache and queue status visibility via the health and status endpoints.

Fixes​

  • [Model Zoo] Fixed a crash during FLUX model inference caused by a dtype mismatch.

Version 0.20.0 (May 22nd, 2026)​

This release adds reinforcement learning training, a model cache for faster repeat inference, and upgrades the core ML stack.

New Features​

  • [Experiments] Training with reinforcement learning (RL) and LLM reward.
  • [Experiments] You can now select which columns to include or exclude for tabular problems, with automatic column suggestions.
  • [Platform] LRU model cache with a configurable VRAM budget, speeding up repeat inference.
  • [Platform] Core ML stack upgraded to Transformers v5, vLLM 0.19.1, and PyTorch 2.10, with a CUDA 12.8 toolchain across all services.
  • [Platform] Tabh2o moved to a dedicated on-prem HTTP service, improving air-gapped support.
  • [Platform] Improved Content Security Policy (CSP) headers.

Fixes​

  • [Datasets] Numerical columns are no longer misclassified as categorical during tabular data loading.
  • [Datasets] Fixed an issue that prevented new dataset uploads.
  • [Projects] Fixed an issue where renaming a project didn't work correctly.
  • [UI] Fixed a theme flicker and a hydration mismatch that could occur on page load.
  • [Platform] Worker components can now access the HF_TOKEN secret for gated Hugging Face model downloads.
  • [Platform] Swagger UI assets are now self-hosted, with HTTPS-aware security headers behind TLS-terminating ingresses.
  • [Platform] Fixed Model Zoo pod startup and Docker build failures.
  • [Platform] Fresh-namespace installs no longer show a temporary login error while the UI finishes starting up.
Feedback