Top leaderboard ranks · read
H2O.ai benchmarks
Generative AI · Decision models
#1 of 159 on JevBench Capability
Decision models get a situation and typed questions and must return calibrated answers, scored on intelligence, calibration, speed and cost. A 31B model with open weights.
Capability, latency and cost, side by side
CompositeCapability top 1010 · sorted by Capability · higher is better
Hatched: the board's Jev reference, a closed API listed but not ranked. Capability, the board's headline score, ranks the 159 inside its cost and latency caps. On the board's official Composite for open-weights systems H2O-Lightning-31B is #3 of 220. Latency includes network: for self-hosted models the board estimates it as 2 × its measured GPU time + 150 ms. Cost is the board's estimate. Benchmark Heaven, board v1.7.48.
Generative AI · Image decision models
#1 of 54 on ImageJevBench Capability
Image decision models answer typed questions about a picture and must return calibrated answers, scored on intelligence, calibration, speed and cost. An open-weights model.
Capability, latency and cost, side by side
CompositeCapability top 1010 · sorted by Capability · higher is better
Capability, the board's headline score, ranks the 54 Jev-class systems: those inside 2 × the cost and median latency of the board's Jev reference. On the board's official Composite H2O-Lightning-31B is #5 of 63. Latency includes network: for self-hosted models the board estimates it as 2 × its measured GPU time + 150 ms. A self-hosted system's cost is its measured GPU time at the board's hourly rate. Benchmark Heaven, board v0.3.0.
Predictive AI · Time-series forecasting
#1 of 32 on fev-bench
Zero-shot forecasting on real time series across many domains and frequencies, scored by probabilistic accuracy against a seasonal-naive baseline. 108.8M parameters.
Win rate · top 8 of 32
higher is better
Default leaderboard view. With the closed-API models included it is #2 of 36. TimeH2O-v1 has a non-commercial licence.
Agentic AI · Future prediction
25 FutureX medals, 8 of them first places
Agents predict real-world events before they resolve, with a new board every week, scored once the answers are known.
Overall score, weekly board · top 8 of 24
higher is better
| # | Model | Framework | Organization | Score |
|---|---|---|---|---|
| 1 | sonnet-5.5 - v2 | H2O AI Super Agent v1.82 | h2o.ai | 52.16 |
| 2 | sonnet-5.5 - 3x | H2O AI Super Agent v1.82 | h2o.ai | 46.60 |
| 3 | GLM 5.2 | arce #1 | Indiv. AI Researcher | 46.05 |
| 4 | grok-4.7+sonnet-5.5 - v4 | H2O AI Super Agent v1.82 | h2o.ai | 45.56 |
| 5 | sonnet-5-5 | H2O AI Super Agent v1.82 | h2o.ai | 45.49 |
| 6 | grok-4.7+sonnet-5.5 - v1 | H2O AI Super Agent v1.82 | h2o.ai | 45.49 |
| 7 | GLM 5.2 | arce #2 | Indiv. AI Researcher | 45.00 |
| 8 | blue-fish | blue-fish | Google DeepMind | 43.88 |
Best of 8 H2O submissions. The agent runs on third-party base LLMs; medals count every H2O entry.
Predictive AI · Tabular foundation models
#5 of 13 on TALENT · #7 of 39 on TabArena
Two public benchmark suites of tabular classification and regression datasets, comparing tabular foundation models, AutoML and tuned gradient boosting.
Try it now
Run TabH2O on your own table in the browser, then get an API key for the REST endpoint.
Open the playgroundTALENT · average rank across datasets · top 8 of 13
lower is better
TabArena · average rank across datasets · top 8 of 39
lower is better
Agentic AI · Social forecasting
#1 of 24 on Social Simulation Arena
H2O AI Super Agent ranks #1 of 24 on Social Simulation Arena's Season 0 pairwise win rate
Entrants forecast the next public release of poll, sentiment and attention series. Forecasts are locked before the answer exists and scored against the published number.
Pairwise win rate · top 8 of 24
higher is better
The site's default view: Season 0 pairwise win rate, 43 questions locked from 30 Sep 2026, one row per method. By arena score over every resolved number question Super Agent is #3 of 25 and Marina #4; by arena score on the Season 0 main board Super Agent is #5 of 26 and Marina #7. Entrants answer different numbers of questions.