H2O Cascade (formerly H2O Enterprise LLM Studio) is a secure model factory. It generates training data from a frontier model, fine-tunes a small model with LoRA, QLoRA, or full training, evaluates it, and serves it through Model Zoo, inside your infrastructure. It covers language, vision, multimodal, and tabular.

Can you fine-tune an LLM without labeled data?

Yes, with H2O Cascade. Reinforcement learning with an LLM as the judge scores each answer from 1 to 10, and the score is the reward. GRPO runs underneath, and no labeled outputs are required. Cascade can also generate training rows from a frontier model or any OpenAI-compatible endpoint.

Which LLMs does Enterprise h2oGPTe support?

OpenAI, Anthropic, Amazon Bedrock, Azure OpenAI, Databricks, self-hosted vLLM endpoints, Cascade-tuned models, and open-weight Danube and Mississippi. Intelligent model routing and per-user cost controls are built in.

Are H2O.ai's models open weights?

Danube language models and Mississippi vision-language models are released under Apache 2.0. tabH2O weights are not released; access is via the hosted API, an on-prem service, or inside H2O Cascade.

Generative AI on your data, your models, your terms.

Answer from your documents with citations. Fine-tune small models that beat frontier prices. Run any LLM from any host, or bring your own, all inside your perimeter and available to the H2O AI Super Agent.

Request a demo How the Super Agent uses it

Part of the Trusted Sovereign AI Platform.

Key takeaways

The H2O Generative AI Platform has two lanes: Enterprise h2oGPTe consumes models for Agentic RAG and Document AI, and H2O Cascade produces fine-tuned models you own.
Run any LLM from any host, OpenAI, Anthropic, Bedrock, Azure, Databricks, vLLM, or your own weights, with cost controls per user and team.
AT&T fine-tuned small language models on H2O.ai and reached 91% accuracy at 35% of the prior cost.
Danube language models and Mississippi vision-language models ship as open weights under Apache 2.0; Mississippi has 2.4M+ monthly downloads.
H2O Cascade can fine-tune without labeled data, using reinforcement learning with an LLM as the judge.

What is Generative AI

Generative AI reads your documents and generates answers, extractions, and actions from them.

The pilot always works. Then the bill, the data, and the labels arrive, and generative programs break in four predictable places on the way to production.

01UnaffordablePilots are cheap. Per-token pricing against a frontier model scales linearly with usage, forever, with no path to a lower unit cost.
02UnprotectedBanks, telcos, healthcare, and the public sector cannot send customer records to a third-party API. Some cannot reach the internet at all.
03UnlabeledFine-tuning projects die on the six months of annotation in front of them, a dataset nobody has the budget to label by hand.
04UnrepeatableLabel in one tool, train in a notebook, track in a third, hand off a checkpoint. Every seam loses reproducibility and the audit trail.

The record

91%accuracy at 35% of prior cost (AT&T SLM)

2.4M+monthly Mississippi VLM downloads

25,000access-center hours returned (health)

Apache 2.0Danube and Mississippi open weights

What the H2O Generative AI Platform does

From documents to answers, and from frontier models to models you own.

Two lanes. Enterprise h2oGPTe consumes models and turns knowledge into governed answers and actions. H2O Cascade produces models, distilling frontier capability into small ones that run where your data lives.

Agentic RAG with citationsAsk across contracts, policies, tickets, and scans. Every answer traces to the document and passage it came from, with multi-step retrieval when one lookup is not enough.
Document AI in the chatOCR and layout understanding for faxes, scans, forms, and tables. Extraction that once needed a trained model now runs on Mississippi vision-language models.
Any model, any hostRoute to OpenAI, Anthropic, Bedrock, Azure, Databricks, or a vLLM endpoint you run. Swap models as the market moves, with cost controls per user and team.
A secure model factoryGenerate training data from a frontier model, fine-tune a small model, evaluate it, and serve it. One governed workflow, one API, inside your perimeter.
Fine-tune without labelsReinforcement learning with an LLM as the judge: the judge scores each answer, the score is the reward. No labeled outputs, and AutoML picks the next run.
Open weights you can inspectDanube language models and Mississippi vision-language models, released under Apache 2.0. Open weights you can read, run offline, and build your own fine-tunes on.

What's in it

The tools in the H2O Generative AI Platform.

Enterprise h2oGPTeThe enterprise GenAI platform that ships the Super Agent.Explore H2O CascadeFrontier capability, flowing down into a model you own.Explore H2O-Danube and H2OVL MississippiSmall open-weight models you own.Explore H2O Document AIBeyond templates and RPA. Models that learn the document.Explore

How you use it

The Super Agent reads, reasons, and acts on your documents.

Agentic RAG is how the H2O AI Super Agent grounds every plan in your own knowledge before it acts. It retrieves, cites, cross-checks, then calls the tools that do the work, from a Cascade fine-tune to a Document AI extraction or a REST call into your systems.

Grounded by defaultEvery answer carries a citation to the page and the passage it came from. No citation, no answer.

Model-agnostic by designA frontier model for the hard reasoning step, a Cascade-tuned small model for the high-volume step. The agent routes, you set the policy.

Inside your perimeterDocuments, embeddings, and fine-tuned weights stay in your own environment, on Managed Cloud or Hybrid Cloud.

Where it fits

Cascade produces. h2oGPTe consumes. The Super Agent puts both to work.

Documents and knowledge flow into h2oGPTe for Agentic RAG and Document AI. Cascade turns your data into fine-tuned models that h2oGPTe serves alongside any external LLM. Highlighted boxes are the H2O Generative AI Platform.

Deliberate boundary: the Feature Store does not feed Cascade, and MLOps does not deploy to h2oGPTe. The lanes meet in the H2O AI Super Agent, not in the plumbing.

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

Proof

Small models, large results.

TelecommunicationsAT&TFine-tuned small language models reached 91% accuracy at 35% of the prior cost, in AT&T's own environment.Read the case study

HealthcareA US academic health systemH2O Document AI reads and routes 1.4M faxes a year, returning 25,000 access-center hours and 5,000 clinician hours. Published result, not a current customer.Read the case study

From LLM Studio to Cascade

Three names, one line of continuous engineering.

Each rename marked a step up in scope. The product outgrew "LLM Studio" once it covered vision, multimodal, and tabular, with data generation, AutoML, RL, and serving attached.

2023H2O LLM Studio shipsopen-source, no-code LLM fine-tuning. The h2oGPT paper follows, and Enterprise h2oGPTe after it.
2024Danube language models and Mississippi vision-language models released as open weights under Apache 2.0.
2025H2O Enterprise LLM Studiofine-tuning as a service, validated on Dell infrastructure.
2026H2O Cascadethe secure model factory. Language, vision, multimodal, and tabular in one workflow.

Part of the H2O AI Platform

Four more capabilities, one platform.

Agentic AIAgentic AI on the Trusted Sovereign AI Platform: the H2O AI Super Agent makes, orchestrates, and governs agents over MCP and A2A, with a full audit trail.Explore ObservabilityAI observability on the H2O AI Platform: explainability with AutoDoc, agent and RAG evaluation, drift and fairness monitoring, and one audit trail.Explore PlatformThe H2O AI Platform unifies predictive, generative, and agentic AI with observability, orchestrated by the H2O AI Super Agent. Deploy, govern, and own it.Explore Predictive AIPredictive AI on the Trusted Sovereign AI Platform: Driverless AI with AutoDoc and AutoMLI, tabH2O, Feature Store, MLOps, driven by the Super Agent.Explore

Videos

Watch Generative AI in action

VideoLLM - Detect AI Generated TextWatch

VideoH2O Generative AI Ecosystem OverviewWatch

VideoBuilding, Evaluating, and Optimizing your RAG App for Production | Jerry LiuWatch

FAQ

Questions people ask before they call us.

Agentic RAG is retrieval-augmented generation where retrieval is a reasoning step, not a single lookup. The agent retrieves, cites the page and passage, cross-checks, and then acts by calling tools. Enterprise h2oGPTe provides Agentic RAG with citations across documents, tickets, and scans.

Next step

Bring one high-volume workflow. Leave with a model you own.

Claims intake, KYC files, contact-center transcripts. We generate the training data, tune a small model, and benchmark it against what you pay a frontier vendor today.

Request a demo Talk to H2O