H2O.ai Releases H2O-Lightning-4B, an Open-Weight Decision Model
H2O.ai has released H2O-Lightning-4B, an Apache-2.0 open-weight decision model fine-tuned from Qwen3.5-4B that achieved a top composite score of 72.5 on the JevBench leaderboard. Unlike traditional text-generating LLMs, it outputs calibrated probabilities for typed questions in a single forward pass without generating text tokens. This release enables ultra-fast, low-cost automated decision-making for enterprise workflows while allowing full local deployment without API fees or data privacy risks. It offers developers a performant open-source alternative to proprietary decision-focused AI models like Jev. Built on Qwen3.5-4B and served via stock vLLM with a lightweight shim, the model processes decisions in approximately 30 milliseconds on an NVIDIA H100 GPU. H2O.ai also announced that larger 12B and 31B open-weight variants are currently in development.
## BACKGROUND
Standard Large Language Models generate responses token-by-token, which introduces output parsing overhead and high latency for structured classification tasks. Decision models operate as rapid evaluators that consume context alongside predefined rubrics to return probability distributions directly over fixed choices in a single step. JevBench is a dedicated benchmark created to measure the intelligence, calibration, speed, and cost of these non-generative decision models.