H2O-Lightning-4B v1.4.1v1.7 · 10 s rule
Data: JevBench v1.7.13, published 2026-10-28 · method v1.7 (10 s rule), measured under the rule
Capability 77.0 · Not ranked in the current release
Measured on the 27 Oct 2026 draw, which is not yet equated to the reference scale. This draw differs strongly from earlier draws (72% multilingual items). On the 300 public items every draw shares, this version scores 57.2 Intelligence vs 61.1 for H2O-Lightning-4B v1.1; its higher overall Intelligence comes from this draw's sealed items. Shown unranked until the draw is equated.
Last measured: 2026-10-27 · measurement release: v1.7.13
Apache-2.0 (HF metadata and LICENSE), public, not gated
intelligence
64.4
calibration
89.7
speed
93.9
cost
65.3
Composite (secondary)
76.0 · Not ranked on this board
Cost and measurement conditions
$0.014 per 1,000 decisions (estimated)
ESTIMATE (*est.), self-hosted open weights, no public tariff: base Qwen/Qwen3.5-4B, deepinfra:Qwen/Qwen3.5-4B at USD 0.03/0.15 per million input/output tokens (frozen 25 Sep 2026 snapshot; listing later deprecated, frozen reference kept; pricing module sha256 645fd82eac728d597048bec0995ecd3ac896566e269fb6c9a1fdb84c171a9ac4). Input is billed on the prompt tokens actually processed by the model: 711,020 tokens after the shipped 32,000-token state trim (chat template and system turn included). The untrimmed rendered count, 944,694 tokens, is disclosed, not priced; 5 items were trimmed at pre-registered positions. Output: 1,500 tokens (one per answered question, as reported). COST FIGURES: USD 0.014370 per 1,000 decisions over all 1,500 items of the draw (the first row on this draw, so no common cost basis applies); 0.22x the official cost cap USD 0.064595. Priced on the untrimmed count it would be USD 0.019044 per 1,000 decisions. Estimated model tariff, not charged usage or a host quote.
Measured p50 latency: 0.020 s (on-GPU, no network, loopback; not directly comparable with API latency, which we time end to end from our measuring host, network included)
Ranking and cap use: 0.19 s (est. incl. network (2x + 150 ms))
Self-hosted model: measured p50 0.020 s on-GPU, no network (timed on our GPU over loopback). Not directly comparable with API latency, which we time end to end from our measuring host, network included. The ranking and latency cap use the same estimate for every self-hosted row: 2 × measured + 150 ms = 0.19 s. Model authors can ask us to time a hosted endpoint instead (on request).
x2 + 0.15 s (assumption, not measured)
Evaluator-owned rented Verda GPU VM, 1x NVIDIA RTX PRO 6000 Blackwell Server Edition (1RTXPRO6000.30V, FIN-03, Finland, EU, pay-as-you-go non-spot, USD 2.649/h all-in), instance id 15b01e61-e76c-4e50-80c0-7aaebf2551d5. Container --network none (cut proven from inside before the sealed input), every credential unset, read-only weights, all 24 model files sha256-verified; driven one request at a time by the frozen JevBench pod driver. The standard self-hosted x2 + 0.15 s latency adjustment applies. Measured 27 Oct 2026 15:22Z on the frozen v1.6-regular-20261027 self-hosted input (sha256 9b2f32ed...); determinism 50/50 bit-identical, 0 argmax flips; VM and volume destroyed afterwards.
Measured latency over 1,500 decisions: p50 0.020 s · p95 0.033 s
0 decisions over 10 s.
A decision that takes longer than 10 seconds (measured time) counts as a wrong answer and as zero calibration for that item: each type's calibration is multiplied by the share of its decisions answered within 10 s. The limit uses the measured time (raw). The displayed adjusted time for self-hosted rows (2 × + 0.15 s) is an assumption and is not used for the limit.