h2oai/h2o-lightning-4b

H2O-Lightning-4B: A Specialized Decision Model

H2O-Lightning-4B, developed by H2O.ai, is a 4.5 billion parameter model based on Qwen/Qwen3.5-4B, designed for efficient and calibrated decision-making. It processes typed questions about records, documents, tickets, and images, providing probabilistic answers for choices, yes/no statements, and ordinal scores. The model runs on unmodified vLLM 0.30.0, with each decision requiring only one forward pass and one output token, optimizing for cost and latency.

Key Capabilities

• Top Performance on JevBench: Achieved the #1 composite score (72.5) on JevBench v1.6.1, surpassing Jev itself (71.5) and larger 12B, 26B, and 31B open models, particularly excelling in calibration and speed.
• Multimodal Decision-Making: Capable of processing both text and image inputs to answer decision questions, supporting up to 4 images per request, each scaled to 1.6 megapixels.
• High Calibration: Designed for probabilities that accurately reflect confidence, a first-class goal in its development.
• Efficient Deployment: Runs with a small standard-library shim, enabling hybrid mode serving alongside the base Qwen3.5-4B model for both decisions and free-text generation from a single vLLM instance.

Good for

• Automated decision-making in applications like processing claims, categorizing tickets, or evaluating policies.
• Use cases requiring highly calibrated probabilistic outputs for classification and scoring tasks.
• Integrating multimodal (text and image) analysis into decision workflows.
• Environments where low latency and cost-effective inference for decision tasks are critical.