v0.0.40

Features

• Tabby: Re-export Tabby from foundationforecast>=0.1.17 for Huawei Paris Noah's Ark Lab's paris-noah/Tabby (arXiv:2609.13956), a ~147M prompted patch transformer with 99 native quantiles (0.01–0.99). Follows the published inference path with a precision option ("bf16" autocast on CUDA by default). Loads with trust_remote_code; weights are CC BY-NC 4.0. See #397.

• TimeH2O: Re-export TimeH2O from foundationforecast>=0.1.16 for H2O.ai's timeh2o-v1 and timeh2o-v1-tiny checkpoints (h2ofm>=0.43.0, Python 3.12+), with clip_nonnegative / snap_integer options matching the published benchmark runs. The checkpoints are gated and released under a non-commercial license. See #397.

Documentation

• Tabby family notebook: Added tabby-family example notebook cross-validating Tabby with its published settings.
• TimeH2O family notebook: Added timeh2o-family example notebook comparing both checkpoints and the benchmark settings.
• Model hub: List Tabby and TimeH2O with license and dependency notes.
• T0 and Timestral notebooks: Removed broken links to the exogenous variables example. See #396.

Dependencies

• FoundationForecast: Require foundationforecast>=0.1.17 (Tabby and TimeH2O wrappers).

---

Full Changelog: v0.0.39...v0.0.40