H2O-Lightning-4B v1.4.1 (image)

Cost and speed evidence

Cost per 1,000 decisions

USD 0.028575

Cost basis

measured GPU seconds x $1.19/GPU-hour

Measured latency · p50 / p95

0.087 s / 0.133 s

Adjusted latency · p50 / p95

0.324 s / 0.415 s

Latency adjustment

v1.4 local latency adjustment

Public accuracy

67.7%

Sealed accuracy

71.1%

Measurement note

Official ImageJev v0.3 S1200+P300 pass 1 on 27 Oct 2026 (1,500 items, the same self-hosted draw as the other GPU rows), paid fast-lane evaluation published under the order's public-visibility setting. Model: h2oai/h2o-lightning-4b at revision cc18ee330e6895c41b9a60f37a0dd122dc4c3461 (v1.4.1), base Qwen/Qwen3.5-4B, bfloat16 weights (no quantization). Served in the author's documented vision configuration: the shipped serve_config.json, the author's two-way profile on one engine with images on (MAX_LEN 40960, GPU_UTIL 0.70), vLLM 0.30.0 with the h2oai per-request-attention overlay (h2oai/vllm @b05cc29a), one RTX PRO 6000 Blackwell, offline after weights and code were hash-verified. The shipped serve_config.json caps images at 1,048,576 pixels, the same per-request cap as the earlier H2O-Lightning-4B image row (v1.1); the vLLM server default of 1,605,632 is overridden per request. Request: one System One choice question with the item question as state, the item image as base64 (data URI only, no image URLs) and the same options; no retries. 1,500 of 1,500 answered, no inference errors. Latency and billed GPU seconds are the wall clock of each loopback request, one request at a time (shim and server overhead included), as for the other H2O-Lightning image rows; other self-hosted rows generally use CUDA-event adapter timing. TypeSafe-20 was not run in this pod session; a follow-up adds those cells. The earlier H2O-Lightning-4B (image) v1.1 row stays on this board as measured.

Published 2026-10-27. Scores and ranks can change in a later release. See the method notes.