AI BriefingH2O.ai Releases H2O-Lightning-4B, Top Open Model on JevBench
Key point
The Apache-2.0 licensed model achieves a composite score of 72.5 on JevBench, surpassing Jev 1.13.
Details
H2O.ai has released H2O-Lightning-4B, an open-weight model licensed under Apache-2.0 designed for decision-style inference. The model processes state and typed questions (such as pick-one, yes-no, or score) to return calibrated probabilities in a single forward pass, eliminating generated tokens for faster and cheaper inference.
Benchmark Performance
On the public JevBench leaderboard, H2O-Lightning-4B achieved a composite score of 72.5, surpassing Jev 1.13 which scored 71.5. This places it as the current top open model on the benchmark.
Technical Specifications
- Base Model: Fine-tuned from Qwen3.5-4B.
- Serving: Compatible with stock vLLM and a small open shim provided in the repository.
- Latency: Approximately 30 ms per decision on an H100 GPU.
- Privacy: Data remains local with no per-call fees.
Future Releases
The company announced that 12B and 31B versions are coming soon. Internal testing suggests these larger models perform better than Jev while maintaining the single-forward-pass architecture.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.