What H2O-Lightning-4B is
H2O.ai released H2O-Lightning-4B on October 7, 2026. It is an open model with 4 billion parameters, built on Qwen3.5-4B, to decide rather than to write:
- A choice between named options, such as the P1, P2 or P3 priority of a ticket.
- Yes or no, with the probability, such as checking whether a claim agrees with its receipt.
- A score on a defined scale, such as urgency on three levels.
On the JevBench ranking, maintained by Benchmark Heaven, version 1.1 ranked first among open models, with the lowest estimated cost per decision among the top entries. It runs on a single GPU, on vLLM.
Who it is for
- Contact centers: prioritize tickets and flag what needs a person.
- Claim and refund review: check the claim against the rule before a person sees it.
- Agents that decide many times per task, where each step adds to the cost.
Why through Rota Nacional
- It runs on Rota's workers. An organization that did not accept sending data to external providers can use it too.
- By default, CPF and CNPJ numbers, email addresses, phone numbers and names leave the text before the model.
- Same API, same billing, with usage records and credit in reais.
How it works
You describe the case
The decisions, the answer options and the volume.
We test with your samples
H2O.ai has not published figures for Portuguese. So we measure the quality in Brazilian Portuguese with your own cases before production.
We connect it to your access
We bring the model up on the workers, connect it to your key and agree on a usage price.
Good to know
- Apache-2.0 license, according to H2O.ai, checked on October 17, 2026.
- The first place on JevBench belongs to version 1.1. The public version is now 1.2.2, which also reads images.
- It does not explain the decision. To write text, the same GPU uses the base model.
- On-demand deployment: the model goes live when an organization asks for it.