H2O-Lightning-4B (image)

Cost and speed evidence

Cost per 1,000 decisions

USD 0.078213

Cost basis

measured GPU seconds x $1.19/GPU-hour

Measured latency · p50 / p95

0.210 s / 0.316 s

Adjusted latency · p50 / p95

0.570 s / 0.781 s

Latency adjustment

v1.4 local latency adjustment

Public accuracy

61.7%

Sealed accuracy

63.5%

Measurement note

Official ImageJev v0.3 S1200+P300 pass 1 on 12 Oct 2026 (no new draw: the same 1,500 items as every other self-hosted row), paid fast-lane evaluation. Model: h2oai/h2o-lightning-4b at tag image (revision 9e9928c7). Served as the model card documents: vLLM 0.30.0 (official image, CUDA 13.0 build; the card names the cu129 wheels of the same release, whose official image could not start vLLM) with the author's unmodified standard-library shim and serve_config.json (image answers at the author's temperature 0.65, images scaled to at most 1,048,576 pixels), offline after the weights were staged and hash-verified. Request: the same System One shape as the Standard One rows (the item question as state, one choice question with the item image). Latency and GPU time are the wall clock of each loopback request, one request at a time. All 1,500 items answered (0 errors). The first request of the run includes the server warm-up (about 25 s) and is counted in GPU time. No retries. Timing basis differs from the other self-hosted v0.3 rows: their GPU time is CUDA-event adapter time, this row's is the loopback HTTP wall clock (shim and server overhead included). This makes its cost higher than a pure GPU-time basis would. Without the warm-up request the cost would be about USD 0.071 per 1,000, still above the Capability cost cap; a pure GPU-time basis could fall under the cap. The row is published on the measured basis.

Published 2026-10-12. Scores and ranks can change in a later release. See the method notes.