Open source, self-hosted. There is no hosted API. You download the weights and run them on your own hardware. How open-source models are listed.
What H2O-Lightning-31B is
H2O-Lightning-31B is a decision model from H2O.ai, published on Hugging Face on 20 October 2026. It is the larger sibling of H2O-Lightning-4B, built on Google’s Gemma-4-31B-it instead of Qwen3.5-4B. It answers Choice, Score and Noul questions about a record, such as a document, a ticket or a conversation, and about images sent with it. Every answer carries a probability for each option.
Each decision is one forward pass and one output token, so the cost is input tokens only.
Licence and access
The weights are gated on Hugging Face under the H2O Model Non-Commercial Licence v1.1. Access is approved automatically once you accept it. The licence allows non-commercial, non-production use only, and the Jev Decision Index notes that evaluation and benchmarking are allowed. That is a stricter licence than the 4B model’s Apache 2.0. H2O has not published a hosted API, and OpenRouter does not list it, so you cannot call it from a paid product.
How you run it
H2O’s recommended setup is its own build of vLLM 0.30.0 with per-request attention, on one RTX PRO 6000 with 96 GB, behind a Python shim that serves POST /v1/systemone, the request shape Jev uses. That build reads short requests in both directions. Stock vLLM on one H100 80GB also works, reading one way only, and scores lower: 65.29 against 66.69 on the Index’s public suite, in H2O’s figures. A state of up to 126,976 tokens is read whole, and a longer one is refused.
What the benchmarks show
On Benchmark Heaven’s JevBench, measured on 22 October 2026, it has an Intelligence score of 80.2, the highest among open-weights entries, and a Capability Score of 85.2. Its Composite of 73.4 is second on the open board, behind H2O-Lightning-4B. The Cost axis, 50.6, holds the composite down, because the board estimates self-hosting cost from a reference price for the base model.
On the Jev Decision Index 0.4.1, generated on 22 October 2026, it scores 66.58 on the Full score, 16th of 229 rows, against Jev 1.14’s 60.24. Its public score is 66.69, from a run H2O submitted. Its calibration error is 0.0842, slightly worse than Jev 1.14’s 0.0799. H2O’s card notes that its score on the Index’s new-domain tests is the lowest among the entries ranked near it.
What it is not for
It is not a chat model and does not explain its answers. H2O says it was evaluated mostly in English and takes no video. A question can have at most 255 options. Retrieval is its weak area on the Index, where its calibration error for that area is 0.266.
Specifications
| System One Score | 96.5 from 2 of 3 boardsJev Decision Index 93.1, JevBench 100. How the score works |
| Question types | ChoiceScoreNoul |
| Max Choice options | 255 |
| Score levels | Not documented |
| Questions per call | Not documented |
| Total context | 131,072 tokens |
| State budget | 126,976 tokens |
| Rate limit | None. It runs on your own hardware. |
| Endpoint | POST /v1/systemone on your own server, through lightning_31b_shim.py in front of vLLM |
| Licence | H2O Model Non-Commercial Licence v1.1 |
| SDKs |
In the card's Decision Index profile the server runs a 131,072-token context and reads a state of up to 126,976 tokens whole; a longer one gets HTTP 422 and is never cut. A question can have up to 255 options. The recommended setup is H2O's own build of vLLM 0.30.0 on one RTX PRO 6000 (96 GB), which reads short requests in both directions; stock vLLM on one H100 80GB works one way only and scores lower. The download is about 63 GB. Evaluated mostly in English, no video input.
Versions
- h2oai/h2o-lightning-31b v1.0.12, The release the model card describes (checkpoint step 22276). A full fine-tune of google/gemma-4-31B-it; the Jev Decision Index lists 31.3B served parameters. JevBench's row is labelled v1.0.7, and H2O's card says the board ran the same weights as this release. Release notes
Use cases
What people use H2O-Lightning-31B for, one page per pattern.
Support inbox triage with System One models
Send a support ticket to Jev once with every question attached. Category comes back as a selected label, severity and frustration as numbers on scales you wrote, refund intent as a probability. Your code reads those values and decides what happens to the ticket.
Compliance verification with System One models
A compliance review is a fixed checklist run against a changing document. Jev answers the whole checklist in one request, returning a probability, a label or a rated level per item, so code can clear the clear-cut findings and send the borderline ones to a person.
Confidence-gated actions with System One models
Jev returns a confidence value from 0 to 1 alongside every Choice and Score answer. Your code treats it as a separate axis: act automatically when it's high, confirm or flag when it's middling, hand the decision to a person when it's low. Riskier actions get higher bars.
Benchmarks
Published scores, each linked to its source. A row marked vendor-run was measured by the vendor itself, so read it as one side's result. The compare hub sets these beside the other models.
| Benchmark | Metric | Score | Run by | Date |
|---|---|---|---|---|
| Jev Decision Index 0.4.1 | Full score | 66.58 score | Independent | 22 Oct 2026 |
| Jev Decision Index 0.4.1 | Public score | 66.69 score | vendor-run | 22 Oct 2026 |
| Jev Decision Index 0.4.1 | Expected calibration error | 0.0842 ECE | Independent | 22 Oct 2026 |
| JevBench 1.7.47 | Composite Score | 73.4 score | Independent | 22 Oct 2026 |
| JevBench 1.7.47 | Capability Score | 85.2 score | Independent | 22 Oct 2026 |
- Jev Decision Index, Full score: Row: H2O Lightning 31B, 16th of 229 rows by this score. Counts 60% private same-skill tests and 40% private new-domain tests; public benchmarks no longer count. 0 is random guessing, 100 is perfect. The Index ran the private tests on one RTX PRO 6000 with H2O's serving in FP8. Jev 1.14 scores 60.24.
- Jev Decision Index, Public score: Row: H2O Lightning 31B. H2O submitted the public run, and the Index rescored it with its own kit. Jev 1.14 scores 60.7.
- Jev Decision Index, Expected calibration error: Row: H2O Lightning 31B. Lower is better. Jev 1.14 scores 0.0799.
- JevBench, Composite Score: Row: H2O-Lightning-31B v1.0.7, measured 2026-10-22 in board release v1.6.39. Second on the open-weights board, behind H2O-Lightning-4B. Intelligence 80.2, Calibration 90.2, Speed 89.8, Cost 50.6.
- JevBench, Capability Score: Row: H2O-Lightning-31B v1.0.7. The average of Intelligence and Calibration, with speed and cost left out. Second among open-weights models on this view.