H2O-Lightning-4B v1.3.1 has public weights, a two-way-reading serving shim and a runnable demo.
H2O.ai updates its 4B decision model with two-way reading
2026-10-17Frontier watchTRIAL
H2O.ai released H2O-Lightning-4B v1.3.1. The 4B model returns choice, yes/no or score probabilities in one forward pass and one output token. The update defaults to two-way reading, handles text and images, and uses a small shim in front of unmodified vLLM. Weights and a live demo are public.
Availability: Apache-2.0 weights, shim and Hugging Face demo
What this means for you
For routing, review and policy decisions, one forward pass can separate latency and cost from generative models. The top leaderboard entry uses older v1.1 weights; v1.3.1 shows no significant public-set gain and a small vision regression, so the rank does not transfer directly to the new checkpoint.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Model | 4B model based on Qwen3.5-4B; text and image decisions |
|---|---|
| Interface | choice, yes/no and score probability outputs; one forward pass and one output token |
| Availability | Apache-2.0 weights, shim and Hugging Face demo |
| Evidence | v1.1 ranks first on the external JevBench open board; v1.3.1 public-set differences are not statistically significant and private board was not rerun |
| Time Basis | v1.3.1 weights committed 2026-10-17T04:41:17Z, inside the previous window but missed |
Technical impact 3
Engineering value 5
Adoption signal 2
Maturity 2
Verification cost 2
Risk 2