Last updated 16 Nov 2026. Sources last checked 16 Nov 2026.
H2O-Lightning-4B
H2O.ai · Released 7 Oct 2026
Learn how to get itDiscuss on X
A 4-billion-parameter open decision model, built on Qwen3.5-4B, that returns a probability for every bounded answer in a single forward pass. The v1.2.2 public release adds image input, with up to four images per request.
- Maker
- H2O.ai
- Release date
- 7 Oct 2026
- First seen
- 7 Oct 2026
- Access
- Apache-2.0 weights on Hugging Face (h2oai/h2o-lightning-4b); runs on unmodified vLLM 0.30.0 with a small Python shim.
- Pricing (as publicly stated)
- Open weights; self-hosted. H2O.ai cited a JevBench board estimate of about $0.021 per 1,000 decisions for the self-hosted v1.1 row, marked as an estimate because open weights have no public price list.
- Modalities
- text, images
- Question types
- choice, noul, score
- Latency (as stated)
- H2O.ai said the board measured a raw median of 29 ms per decision for v1.1, then applied its standard self-hosted allowance before scoring.
- Official announcement
- https://www.h2o.ai/blog/2026/h2o-lightning-jevbench
Benchmark claims
These are claims as published by the maker or in the linked official post. This catalog has not independently rerun every benchmark.
H2O.ai said H2O-Lightning-4B v1.1 ranked first of 104 open-weight systems on the JevBench v1.6.1 official Composite Score, at 72.5 versus 71.5 for Jev 1.13.0, as read on 7 Oct 2026. It also reported a calibration score of 90.0. The company noted that ranks can change as the board is updated, and that the board is run independently by Benchmark Heaven.
What enterprises should know
Cost and pricing
Open weights; self-hosted. H2O.ai cited a JevBench board estimate of about $0.021 per 1,000 decisions for the self-hosted v1.1 row, marked as an estimate because open weights have no public price list. Figures are what the maker stated in the linked source and may have changed.
Lock-in
The official announcement describes distribution as follows: Apache-2.0 weights on Hugging Face (h2oai/h2o-lightning-4b); runs on unmodified vLLM 0.30.0 with a small Python shim. That is the maker's description of how the product is obtained, not an assessment of switching costs.
Security and privacy
The official announcement we sourced did not state a specific security certification or independent privacy audit for this product. If you use a hosted endpoint named in the announcement, inputs are sent to that host under the host's own terms.
Maturity and availability
The maker's official announcement is dated 2026-10-07. Access was described as: Apache-2.0 weights on Hugging Face (h2oai/h2o-lightning-4b); runs on unmodified vLLM 0.30.0 with a small Python shim. H2O.ai said the board measured a raw median of 29 ms per decision for v1.1, then applied its standard self-hosted allowance before scoring.
Compliance notes
The official announcement we sourced did not state a specific regulatory approval or compliance certification for this product.
What consumers should know
Cost and pricing
The announcement describes a way to obtain or run the model (including, where stated, open weights or self-hosting). Hosted prices, if any, are as follows: Open weights; self-hosted. H2O.ai cited a JevBench board estimate of about $0.021 per 1,000 decisions for the self-hosted v1.1 row, marked as an estimate because open weights have no public price list.
Lock-in
The announcement describes access as: Apache-2.0 weights on Hugging Face (h2oai/h2o-lightning-4b); runs on unmodified vLLM 0.30.0 with a small Python shim. Self-hosting, where it is offered, means you are not limited to the maker's hosted endpoint, but you still depend on the published weights, license, and your own hardware.
Security and privacy
The announcement describes local or self-hosted use. What happens to your data then depends on where you run the software. The announcement we sourced did not state a product-specific consumer privacy audit.
Maturity and availability
The official announcement is dated 2026-10-07. Availability was described as: Apache-2.0 weights on Hugging Face (h2oai/h2o-lightning-4b); runs on unmodified vLLM 0.30.0 with a small Python shim. H2O.ai said the board measured a raw median of 29 ms per decision for v1.1, then applied its standard self-hosted allowance before scoring. This summary does not say whether the product is suitable for a particular consumer task.
Compliance notes
The official announcement we sourced did not state consumer-protection certifications specific to this product.
Figures are as stated in the linked source on the date shown. Sources last checked 16 Nov 2026. Names and marks belong to their owners. This page is an independent summary, not a product page from H2O.ai.