Zero-shot probabilistic time-series foundation model in 146K parameters, INT8-deployable on embedded hardware
-
Updated
Sep 12, 2026 - Python
Zero-shot probabilistic time-series foundation model in 146K parameters, INT8-deployable on embedded hardware
First Apache-2.0 Turkish-vertical TSFM fine-tune: Chronos-2 + LoRA r=32 + V4 router (MASE-aware adoption) + bucket isotonic calibration. GIFT-Eval submission 2026-05-27.
Run, compare and inspect time-series foundation models (Timestral, Chronos-2, TimesFM, t0-gamma, TiRex-2) on your own machine. Unofficial, local-first.
Tiny (4.73M) zero-shot time-series foundation model. 19 MB checkpoint, ~6 ms CPU inference, competitive on GIFT-Eval.
Launch apps, search files, and run commands instantly with a native macOS hotkey launcher that stays under 100 MB RAM.
Modular PyTorch time-series foundation model (14.8M–16.3M) pretrained on GiftEvalPretrain across 6 controlled ablations and evaluated zero-shot across all 97 GIFT-Eval tasks.
To associate your repository with the gift-eval topic, visit your repo's landing page and select "manage topics."