H2O-Danube and
H2OVL Mississippi
Small open-weight language and vision-language models, trained from scratch by H2O AI under Apache 2.0, built to be fine-tuned and to run inside your perimeter, even offline.
- H2O-Danube small language models (500 million to 4 billion parameters) and H2OVL Mississippi vision-language models (0.8 and 2 billion) are open-weight models trained from scratch by H2O AI and published under Apache 2.0 on Hugging Face.
- Danube is built to be fine-tuned: Danube3-500M is the default backbone in H2O Cascade, so a frontier model can label the data once and a small Danube model serves the task from then on.
- Mississippi reads forms, receipts, statements, charts and tables into structured data with intelligent character recognition (ICR), so routine pages are read inside your perimeter.
- Both families run on a laptop, an edge device, or a server in an air-gapped data center. Danube also runs on a phone for inference.
- The Mississippi models had more than 2.4 million monthly downloads, according to H2O AI in July 2026.
Two families, one reason: a model that stays inside your boundary.
A model behind someone else's API sends every call outside your perimeter and bills you for each one. Open weights answer both. Small ones also answer where it can run.
- Danube language modelsDanube3 ships in 500M and 4B parameters (July 2024), trained from scratch by H2O AI, with base and chat versions on Hugging Face.
- Mississippi vision models0.8B and 2B vision-language models (October 2024), built on Danube with an InternViT-300M encoder, that read documents, charts and tables.
- Runs where the data isSmall enough for a laptop, an edge device, or a GPU node in an air-gapped data center, and Danube runs on a phone for inference.
- Cascade's default backboneDanube3-500M is the default backbone in H2O Cascade: a frontier model generates the training rows, Danube learns the task, your GPUs serve it.
- Reads pages in your perimeterMississippi reads forms, receipts, statements, charts and tables into structured data with ICR. Keep the frontier model for the hard pages.
- Apache 2.0 weightsUse, modify, fine-tune and ship commercially under Apache 2.0, with no per-token bill and methods published in three technical reports.
From download to a model you own.
- 01DownloadPull the weights from Hugging Face: base and chat versions, GGUF for llama.cpp, and ONNX for Danube3-500M, under Apache 2.0.
- 02RunOn a laptop, an edge device, or a GPU node inside your air gap. Serve through Enterprise h2oGPTe, vLLM, llama.cpp, or your own runtime.
- 03Fine-tuneIn H2O Cascade, a frontier model generates the training data, then you tune Danube with LoRA, or with RL and a judge, and evaluate it.
- 04Own itThe resulting model is yours to run where you choose, with no per-token bill and no vendor terms that can change under you.
The open foundation of the generative lane.
Danube is the default backbone of H2O Cascade, and Enterprise h2oGPTe serves both families beside external LLMs. Highlighted: the open models and the tool that fine-tunes them.
Shipped, and safe to claim
Use the frontier model where it earns its cost.
Enterprise h2oGPTe, which ships the H2O AI Super Agent, can route each call to a model on latency, cost and accuracy. A guardrail check, a page read or a classification does not need a frontier model, and on Danube or Mississippi it does not leave the building.
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.
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.
Questions people ask before they call us.
H2O Cascade uses Danube3-500M as its default fine-tuning backbone, so a frontier model can label the data once and a small Danube model can serve the task from then on. Enterprise h2oGPTe serves Danube and Mississippi beside external LLMs and can route each call on latency, cost and accuracy. Cascade also tunes other open models, so Danube is a default, not a requirement.
Bring a document type your current character-recognition engine gets wrong.
We will run it through Mississippi inside your environment and show you the extraction, the cost, and where the data did not go.