H2O-Lightning-4B v1.1 prints JevBench composite 72.5 against Jev at 71.5

Jonathan McKinney’s H2O blog, published October 7, 2026, prints H2O-Lightning-4B v1.1 first on JevBench’s official composite. The table is release v1.6.1, board revision v1.7.17, read that day, and the blog says it ranks 104 open-weight systems. The 4B scores 72.5. Jev 1.13.0 is the unranked reference at 71.5. Quyet-1.0-Large, 31B, is second at 71.4. Note 2 says ranks and scores can change as systems are added.

System	Size	Composite	Intel.	Calib.	Speed	Cost	$ / 1k
H2O-Lightning-4B v1.1	4B	72.5	60.0	90.0	92.6	60.3	$0.021
Jev 1.13.0, reference	n/a	71.5	63.6	90.6	91.5	54.7	$0.032
Quyet-1.0-Large	31B	71.4	73.4	90.0	86.9	50.5	$0.045

The blog says the board’s detail card for this row read 0.65 times Jev’s cost and 0.34 times its median latency. Note 5 separates two clocks. The raw median is 0.029 seconds. The adjusted median is 0.21 seconds, from the board’s “x2 + 0.15 s” allowance on self-hosted rows. Hosted APIs are not adjusted. On the capability view in the same post, which averages Intelligence and Calibration, the 4B is third among open models, behind two 31B models. The post also names Sage 1.3.0 at 74.0 and Liquid d1 at 73.0 on the API board, and says those calls are a separate ranking.

The weights line says Apache 2.0, on Hugging Face as h2oai/h2o-lightning-4b, from Qwen/Qwen3.5-4B. The serve commands use unmodified vLLM 0.30.0 and a small Python shim. The download pins revision v1.2.1. The blog says the board measured v1.1, the text model, and that v1.2.1 adds image input, up to four images in one request. A request with no image takes the text path that was measured. We did not open the card. We did not run the weights.

A later paragraph in the same post says the decision adapter, on one H100, scores 205 of 231 public JevBench items, the same as the released model, and gives the same top answer on 230 of them. Text such as a 90-token reply comes from the base model, about 0.6 seconds on that GPU. The 205 of 231 figure is that public-item count. It is not the composite 72.5.

The October 8 open of the open-weights list, already written on the JevBench page, records this checkpoint at capability 75.0, capability rank 5, and official composite 1, and does not copy a composite number from that card. That reading is later than revision v1.7.17. It is not a second printing of 72.5.

Question types in the blog’s example are choice, noul, and score, on POST /v1/systemone. The example’s own probabilities, P1 at 0.944 and “affects everyone” at 0.929, are the card’s illustration. They are not a benchmark.

Other open checkpoints are on decision models like Jev.