11st of 2 in Open models
timesfm
Pretrained time-series foundation model for zero-shot forecasting from Google Research
- Stars
- 34.5k
- License
- Apache-2.0
- Last commit
- Sep 2026
- Last release
- Aug 2026
- Language
- Python
9,991 on 24 Mar 202634,475 on 13 Oct 2026
From GitHub's daily star counts.
Checked on GitHub.
Overview
TimesFM is a pretrained decoder-only model for time-series forecasting, installed as a Python package with PyTorch or MLX (Apple silicon) backends. Version 3.0 forecasts univariate and multivariate series with past-only and past-and-future covariates, returning point forecasts plus nine quantiles (0.1 to 0.9). The code is Apache-2.0, but 3.0 weights carry a non-commercial license.
Who it is for: Engineers and researchers forecasting time series without per-task training
Strengths
- Zero-shot forecasting of univariate and multivariate series with past-only and future covariates
- Returns point forecast plus 9 quantiles (0.1 to 0.9) per step
- MLX backend runs 3.0 on Apple silicon without PyTorch, numerically matched to PyTorch
- Includes LoRA fine-tuning example (HF Transformers + PEFT), unit tests and an agent SKILL.md
Weaknesses
- TimesFM 3.0 weights are non-commercial and non-production; commercial use only via Google Cloud services
- Not an officially supported Google product
- Forecasting only; contexts beyond 15,360 points are truncated to the most recent
- README code imports differ between 3.0 examples and the PyPI 2.0.2 note, which is unclear
What it needs
- GPU optional
- Needs PyTorch or MLX, Hugging Face checkpoints
- Models: TimesFM 3.0 (PyTorch checkpoint), TimesFM 2.5 (200M params, 16k context), TimesFM 1.0 and 2.0 (v1 archive)
Head to head
Also in Open models
See all 2| Rank | Project | Score |
|---|---|---|
| 2 | clipOpenAI's reference PyTorch code and weights for contrastive image-text embeddings | 26 out of 100 |