Why was Enterprise LLM Studio renamed H2O Cascade?

The product outgrew its name. It now spans language, vision, multimodal, and tabular, with synthetic data generation, AutoML, reinforcement learning, and serving attached. A cascade is the shape of the workflow: frontier capability flowing down into a small model you own.

Can H2O Cascade fine-tune without labeled data?

Yes. RL fine-tuning with LLM-as-a-Judge uses a judge model to score each answer from 1 to 10 as the reward, with GRPO underneath. Cascade can also generate labeled training rows from a frontier model. This is not RLHF in the strict sense.

Does H2O Cascade deploy to H2O MLOps?

No. Cascade serves through its own Model Zoo, with shared GPUs, hot-swap, and batch inference with priority queueing. Deployment to H2O MLOps and a Python SDK are roadmap items, not shipped.

Does H2O Cascade run air-gapped?

Yes, on Kubernetes/Helm or Docker Compose. In air-gapped installs, data generation from a hosted model requires an internal endpoint. You choose where the model runs.

H2O Cascade

A secure model factory that generates training data, fine-tunes a small model, evaluates it, and serves it, all inside your own infrastructure. Formerly H2O Enterprise LLM Studio.

Request a demo Documentation

The record

4modalities: language, vision, multimodal, tabular

0labeled outputs required, with RL and a judge

35%of prior cost at 91% accuracy, AT&T

2023the LLM Studio line began, open source

Key takeaways

H2O Cascade (formerly H2O Enterprise LLM Studio) is a secure model factory: it generates training data, fine-tunes, evaluates, and serves a specialized model inside your infrastructure.
It covers language, vision, multimodal, and tabular; backbones include Danube, Llama, Qwen3, Gemma, DeepSeek-R1 distills, Mistral, and any Hugging Face model.
RL fine-tuning with LLM-as-a-Judge (GRPO) needs no labeled outputs; AutoML over fine-tuning searches N experiments for you.
AT&T fine-tuned small language models reached 91% accuracy at 35% of the GPT-4 cost on H2O.ai's enterprise fine-tuning platform.
Full REST API and MCP server; no Python SDK today. Serving through Model Zoo, not H2O MLOps.

What it does

The whole workflow, no seams.

Label in one tool, train in a notebook, track in a third, hand a checkpoint to a platform team, and every seam loses reproducibility and the audit trail. Cascade is one workflow.

Generate the dataUse a frontier model, or any OpenAI-compatible endpoint including h2oGPTe, to generate or label training rows, with per-job cost accounting.
Fine-tuneLoRA by default, QLoRA, or full training. Backbones from Danube to Llama, Qwen3, Gemma, and any Hugging Face model, with FSDP and Flash-Attention-2.
Learn without labelsRL fine-tuning with LLM-as-a-Judge: a judge scores each answer 1 to 10 and the score is the reward, GRPO under the hood. The no-labels case, solved.
AutoML for fine-tuningSet a budget of N experiments and Cascade searches configurations for the best. Ask KGM, the Kaggle Grandmaster agent, proposes the next run.
EvaluatePerplexity, BLEU, LLM-as-a-Judge, QA accuracy, AUC, LogLoss, and mAP, scored, compared, and tracked, with a lineage graph for every experiment.
ServeModel Zoo on shared GPUs: hot-swap, bin-packed, batch inference with priority queueing, and a library from vLLM chat to FLUX.1 and Florence-2.

How you use it

The signature workflow.

01Frontier model generatesTraining rows generated or labeled from a large model you point it at.
02Small model learnsLoRA, QLoRA, or full fine-tune. Or RL with a judge when there are no labels.
03Evaluation scoresCompared, tracked, and reproducible. Ask KGM proposes the next run.
04Your endpoint servesModel Zoo, your GPUs, your perimeter. Pay for the model once.

Where it fits

The producer side of the generative lane.

Cascade turns your data into models. h2oGPTe consumes them. Highlighted is Cascade and the open models it builds on.

Deliberate boundaries: the Feature Store does not feed Cascade, and Cascade serves through its own Model Zoo rather than MLOps.

Details

Shipped, and safe to claim

Problem typesCausal LM (with RL), Classification LM, Image Classification, Multimodal (GA), Object Detection (experimental), Tabular via tabH2O

Data generationRows or columns from an approved hosted model or any OpenAI-compatible endpoint; per-job cost accounting

TrainingLoRA (default), QLoRA, full; FSDP, Flash-Attention-2; AutoML over N experiments; Ask KGM

Reinforcement learningGRPO with LLM-as-a-Judge; no labeled outputs required (not RLHF in the strict sense)

EvaluationPerplexity, BLEU, LLM-as-a-Judge, QA accuracy, AUC/LogLoss, mAP

ServingModel Zoo: shared GPUs, hot-swap, bin-packing, batch API with priority queueing

GovernanceDataset, config, metrics, lineage, GPU cost, and owner per experiment; Keycloak/OIDC, RBAC, model allowlist

APIsFull REST API, per-user keys, MCP server. No Python SDK today

DeploymentKubernetes/Helm or Docker Compose; Hybrid Cloud today, Managed Cloud and GovCloud on the roadmap

GPUCompute capability 8.0 or higher; A100/H100 80GB recommended

With the Super Agent

The Super Agent orders the model. Cascade builds it.

Cascade exposes an MCP server, so "fine-tune a classifier on these 80 topics and deploy it" becomes a step the H2O AI Super Agent can take inside a larger plan, with the cost attributed and the lineage recorded.

A fine-tune as a tool callThe agent starts the job, reads the evaluation, and promotes the winner, all inside a larger plan.

Cost attributedGPU hours and generation tokens per job, visible to the team that asked for the run.

Lineage recordedWhich data trained the model in production, and what the run cost, kept on the record.

Run it anywhere

You operate it, or H2O AI operates it for you.

Same platform either way: the same capabilities, the same governance, and the same audit trail. The choice stays open, so you can switch between them as the business changes, without rebuilding what runs on top.

We run itH2O AI Managed Cloud (HAMC)A dedicated, single-tenant cloud that H2O AI provisions, operates, and secures, across fault-tolerant availability zones in the region you choose. Your data never leaves it. Zero-trust access, encrypted in transit and at rest, constantly monitored.AWSAzureGoogle CloudSOC 2 Type 2 · HIPAA / HITECHLearn more You run itH2O AI Hybrid Cloud (HAHC)The full H2O Sovereign AI Platform, entirely within your control, on your own infrastructure or fully air-gapped. You own the environment, the updates, and the compliance posture. Installed and upgraded with Terraform, Helm, or Replicated.Your cloud accountYour data centerAir-gappedLearn more

H2O AI Cloud for Government

The H2O AI Cloud for Government has achieved FedRAMP High Certification, delivering the highest standard of cloud security so federal agencies can harness predictive and generative AI without compromising data sovereignty or compliance.

Learn more View on marketplace

Works with

Videos

Watch H2O Cascade in action

VideoMastering LLM Data Preparation with LLM DataStudio: A Step-by-Step GuideWatch

VideoH2O.ai Launches H2OGPT and LLM Studio: Build Your Own Enterprise Grade ChatbotsWatch

VideoThe Fine Art of Fine-Tuning Large Language Models Talk at H2O World IndiaWatch

FAQ

Questions people ask before they call us.

H2O Cascade, formerly H2O Enterprise LLM Studio, is a secure model factory. It generates training data from a frontier model or any OpenAI-compatible endpoint, fine-tunes a small model with LoRA, QLoRA, or full training, evaluates it, and serves it through Model Zoo on your GPUs. It handles language, vision, multimodal, and tabular.

Next step

Bring the workflow with the biggest per-token bill.

We will generate the data, tune a small model, and benchmark it against your frontier vendor on accuracy and cost.

Request a demo Talk to H2O