H2O Lightning
A family of open-weight decision models: give it a record and a question with fixed answers, and it returns a calibrated probability for every answer in one pass.
- H2O Lightning is a family of open-weight decision models: send a record, with up to four images, and typed questions with fixed answers, and get a calibrated probability for every answer in one forward pass.
- It returns probabilities, not text, so you can threshold: send high-confidence calls straight through, route the middle to a person, and escalate the rest to a larger model.
- The weights are Apache 2.0 on Hugging Face and run with vLLM on one GPU, so the record stays in your environment. H2O Lightning is also available in Enterprise h2oGPTe.
- H2O Lightning decides; it does not write or explain. Use Enterprise h2oGPTe for open-ended answers, and H2O Driverless AI for regulated scoring with reason codes.
Decisions with a probability on every answer.
- Calibrated probabilitiesEvery answer carries a probability, trained so a stated 70 percent is right about 70 percent of the time on decision questions.
- Typed questionsAsk choice, yes or no, and score questions, several per record, with up to 255 answer options each, and read every probability.
- One pass, no textThe model reads the record once and scores each answer at a single token, so output is always valid and never needs parsing.
- Text and imagesA record can include up to four images, such as receipts, screenshots and scanned pages, read with the text. Video is not supported.
- Open weightsDownload Apache 2.0 weights from Hugging Face and run them on your own GPU, with no per-call bill and no vendor endpoint.
- Fail-closed servingThe open shim refuses to answer when the server runs the wrong model or a label is not one token, rather than scoring anyway.
A record in, probabilities out.
- 01ServeRun the weights with vLLM and the open shim from H2O.ai on one GPU, or use the model in Enterprise h2oGPTe, with no training step.
- 02AskSend one request with a record and typed questions. Images go inline, up to four per record, and the questions share one pass.
- 03ReadEach question returns a probability for every answer, a confidence and the top choice, so there is no text to parse or repair.
- 04ActThreshold in your code: send high-confidence calls through, route the middle to a person and escalate the rest to a bigger model.
The fast decision step, beside your large model and your trained models.
H2O Lightning decides; it does not write. Put it in front of a larger model to screen, route and gate, keep Enterprise h2oGPTe for open-ended answers, and use H2O Driverless AI for regulated scoring with reason codes.
What ships today
Questions people ask before they call us.
A prompted model writes its answer as text, which costs output tokens, needs parsing and gives you a confidence you cannot trust. H2O Lightning reads the record once and returns a probability for every answer from a single token, so you can set thresholds and send only the uncertain cases onward. It does not write or reason; use a larger model for that.
Bring a decision you make a thousand times a day.
We will run it on your records inside your environment, and show you the probabilities, the latency on your GPU and where it is not a fit.