AnswersPublished24 sources
What TimeH2O-v1’s Forecasting Benchmark Results Show
H2O.ai reports that its 108.8M parameter TimeH2O v1 achieved a 91.9% win rate and 50.8 skill score across 100 fev bench tasks. The model is designed for zero shot forecasting, and H2O reports GIFT Eval aggregate MASE of 0.654 and CRPS of 0.441 among zero shot models.
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- H2O.ai reports that its 108.8M parameter TimeH2O v1 achieved a 91.9% win rate and 50.8 skill score across 100 fev bench tasks.
- The model is designed for zero shot forecasting, and H2O reports GIFT Eval aggregate MASE of 0.654 and CRPS of 0.441 among zero shot models.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is H2O.ai’s 108-million-parameter TimeH2O-v1 time-series foundation model, released on October 26, and how do its zero-shot forecasting. Article summary: TimeH2O-v1 is H2O.ai’s compact, open-weights time-series foundation model, released on October 26, 2026, for zero-shot forecasting without a separate training run or model for every series.[9][10] H2O.ai reports first pl. Topic tags: general, general web. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fake numbers, clic
H2O.ai released TimeH2O-v1 on October 26, 2026, as a compact time-series foundation model for zero-shot forecasting. The headline is strong: H2O reports a 91.9% win rate and 50.8 skill score across 100 fev-bench tasks. But those numbers describe a particular leaderboard view, not a universal first-place result across every metric or model-access category.1
7
13
What TimeH2O-v1 is
TimeH2O-v1 is an open-weights forecasting model with 108.8 million parameters, often rounded to 108 million. Zero-shot forecasting means applying a pretrained model to forecast a time series without training a separate model for each series.1
6
13
H2O describes fev-bench as evaluating probabilistic forecasting against a seasonal-naive baseline. The benchmark spans 100 tasks; a separate leaderboard description says those tasks draw on 96 real-world datasets across seven domains.1
19
How its fev-bench results compare
In H2O’s listed default win-rate standings, TimeH2O-v1 is first among 32 models. That view lists Timestral-209M at 89.6%, t0-gamma at 88.1%, TimesFM-3 at 82.7%, and Chronos-2 at 76.8%, compared with TimeH2O-v1’s 91.9%.13 A Bernama report also gives the 89.6% and 88.1% figures for Timestral-209M and t0-gamma.
2
The scope of that ranking matters. H2O’s leaderboard says TimeH2O-v1 ranks second among 36 when closed-API models are included. H2O also qualifies its top-line comparison as including non-commercial models.13
6 So “first” applies to the specified leaderboard configuration, not every possible model set.
There is also a meaningful difference between the H2O-listed win rate and another fev-bench listing. TSFM.ai shows TimeH2O-v1 with an 86.14% win rate and a 39.95 MASE skill score, ranking third by that MASE measure; its listing places TimeRSI-1.5 and TimeRSI above it.19 These figures should not be treated as interchangeable with H2O’s 91.9% win rate and 50.8 skill score: the published leaderboard views report different values and rankings. The takeaway is that a benchmark claim needs its metric and leaderboard version attached.
GIFT-Eval: another positive result, with a different metric
H2O reports that TimeH2O-v1 had the lowest aggregate MASE and CRPS among zero-shot models on GIFT-Eval, with scores of 0.654 and 0.441 respectively.6 A separate GIFT-Eval listing shows rounded values of 0.65 for MASE and 0.44 for CRPS, alongside 0.66 and 0.45 for Timestral-209M.
24 These are results on GIFT-Eval, a different benchmark from fev-bench; they should not be used to restate or validate the fev-bench ranking.
Covariates and fine-tuning: what the available evidence supports
The fev-bench description includes tasks with covariates, but the listed fev wrapper for TimeH2O-v1 says its checkpoints forecast each target independently and ignore covariates, converting tasks to univariate series.19
22 That describes the benchmark implementation; it does not establish every capability of the wider product or deployment interface.
The available source material does not document a customer fine-tuning workflow, required infrastructure, or a verified process for turning customer data into domain-specific models. Those details should be confirmed before treating fine-tuning as an established feature.
Training data and benchmark exclusions
TimeH2O-v1’s model card says datasets overlapping with fev-bench, GIFT-Eval, and TIME were excluded from training.4
8 The disclosed training-source examples include AutoGluon’s Chronos datasets, Salesforce’s GiftEvalPretrain and lotsa_data collections, and the Quito corpus. The model card excerpt lists different licenses across those datasets, including Apache-2.0 and CC-BY 4.0; the disclosed list is not necessarily complete.
8
Availability and license
The weights are hosted in H2O.ai’s Hugging Face repository, where users must accept access conditions to download them.10 The repository lists the H2O Model Non-Commercial Licence v1.0. H2O’s product page says production use requires a license from the company, so open weights should not be read as permission for unrestricted commercial deployment.
5
14
H2O also lists a TimeH2O-v1-tiny variant at 5.83 million parameters under the same named non-commercial license.12
14 The provided materials do not establish that it matches v1’s forecasting performance.
Bottom line
TimeH2O-v1 has a strong reported zero-shot result: H2O’s fev-bench view gives it a 91.9% win rate and 50.8 skill score, and H2O reports leading aggregate MASE and CRPS among zero-shot models on GIFT-Eval.1
6 The qualifications are just as important: other leaderboard views show different fev-bench scores and rankings, closed-API inclusion changes its position, and the weights are not licensed for unrestricted commercial production use.
13
14
19
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "What TimeH2O-v1’s Forecasting Benchmark Results Show"?
H2O.ai reports that its 108.8M parameter TimeH2O v1 achieved a 91.9% win rate and 50.8 skill score across 100 fev bench tasks.
What are the key points to validate first?
H2O.ai reports that its 108.8M parameter TimeH2O v1 achieved a 91.9% win rate and 50.8 skill score across 100 fev bench tasks. The model is designed for zero shot forecasting, and H2O reports GIFT Eval aggregate MASE of 0.654 and CRPS of 0.441 among zero shot models.