H2O ships TimeH2O: 108M-param zero-shot forecasting model claims top fev-bench spot
krishnan · x · 2026-11-18
H2O made TimeH2O and TabH2O generally available on H2O AI Cloud.
- TimeH2O: a 108M-parameter time-series foundation model that forecasts zero-shot, with no per-series model building.
- TabH2O: the same zero-shot idea for tabular work like risk scoring and fraud detection.
- H2O reports TimeH2O-v1 ranked first of 32 models on the fev-bench leaderboard as of Oct 23 without training on any fev-bench dataset — but that's a vendor-reported accuracy claim, not evidence of lower inventory costs.
The author's take: if the claims hold, forecasting gets cheap enough for workflows that never justified a dedicated model — an inventory agent can execute every step correctly yet still stock the wrong product if the forecast behind it is wrong.
More from coding & agent
- Astra 6.2 beats Opus at image-gen + coded animation in a Codex-built interactive storybook — gabrielchua · 2026-11-18
- Prompting Claude from an iPhone to submit Slurm jobs to a compute cluster — vxnuaj · 2026-11-18
- Dev runs MCP browser agent on real sites overnight, documents the bugs unit tests miss — JeremyCMorgan · 2026-11-18
- Musk says Grok Bot cut SpaceXSI invoicing from 24 hours to minutes with 5 accounting agents — elonmusk · 2026-11-18
- Users slam OpenAI's dots for lacking custom MCP server support — carlosdponx · 2026-11-18
- User runs 83 AI agents concurrently, nearly 300 combined — gnukeith · 2026-11-18