H2O-Lightning-31B v1.3.4 (@c35ed950)v1.7 · 10 s rule

Data: JevBench v1.8.13, published 2026-11-16 · method v1.7 (10 s rule), measured under the rule

Capability 87.9 · Not ranked in the current release

Added after the v1.8.10 ranking review; it is ranked at the next review.

Last measured: 2026-11-16 · measurement release: v1.8.13

H2O Model Non-Commercial Licence v1.1 (weights; gated on Hugging Face; benchmarking and publishing results are allowed by the licence). Base Gemma 4 under its own licence.

Base model: google/gemma-4-31B-itsource

intelligence

85.0

calibration

90.7

speed

90.3

cost

60.2

Composite (secondary)

79.3 · Not ranked on this board

Cost and measurement conditions

$0.021 per 1,000 decisions (estimated)

GPU-time estimate from v1.8.0: one monthly-rented RTX PRO 6000 96 GB, USD 1.5235/hour; average load 70%. Architecture: dense gemma, 31B active parameters; throughput 10106.516129 tokens/s (architecture-class calibration, efficient-engine prefill). GPU-time cost USD 0.0212260058867847/1,000; old list-price estimate USD 0.0321852518468771/1,000 (google/gemma-4-31b-it at USD 0.09/M input and USD 0.34/M output). Applied min(old list-price estimate, GPU-time cost) = USD 0.0212260058867847/1,000. Estimated, not a bill.

Measured p50 latency: 0.052 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: 0.25 s (est. incl. network (2x + 150 ms))

Self-hosted model: measured p50 0.052 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 = 0.25 s. Model authors can ask us to time a hosted endpoint instead (on request).

x2 + 0.15 s (assumption, not measured)

Evaluator-owned Verda GPU VM (RTX PRO 6000), offline container (--network none; lo only), author's serving stack at model commit c35ed9505f6744770dc0d74bd3f468791a7984af. Unmodified serve.sh --profile two-way, one vLLM 0.30.0 engine with h2oai/vllm overlay a36449a18b4465b4f70fa078f5403032de414de8. Online FP8 W8A8 from BF16 weights, TRITON_ATTN; max model length 131072, state budget 126976, two-way window 8192, selector v2 with automatic one-way fallback for the record/window or summed prompt threshold 4096. One read, readout_orders=1; AR and router off. PyTorch 2.13.0, transformers 5.20.0, torchcodec 0.17.0. Each scored stage is one fresh compile read once (no best-of); determinism is assessed only within the public preparation compile. Standard self-hosted x2 + 0.15 s latency adjustment.

Measured latency over all decisions: p50 0.052 s · p95 0.107 s

— 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.

Model and serving disclosure

Paid public evaluation. Measured on the 27 Oct draw on an evaluator-owned Verda GPU VM (RTX PRO 6000), offline with the author's serving stack at the pinned commit. Each scored stage is one fresh compile read once (no best-of). Failed or unanswered requests count as wrong answers. Second H2O-family model on this draw (after H2O-Lightning-4B v1.4.1); no shared weights. The temperatures and yes/no floor were fitted on v1.3.0's weights from Decision Index 0.3 public per-row reads (the release keeps them). The router head was fitted on HLE-style sets, with JevBench's public 231 items read as held-out. The determinism figure describes the preparation compile only; each scored stage is one fresh compile read once.

Published source