TimesFM 2.5 200M
Release in the TimesFM family · version 2.5 (timesfm-2.5-200m)
Maintained by Google Research1
TimesFM 2.5 is a 200M-parameter time-series forecasting model from Google Research, released in September 2025. Compared with TimesFM 2.0 it has fewer parameters (200M, down from 500M), a context length of up to 16k points, an optional 30M quantile head for continuous quantile forecasts up to a 1k horizon, and no frequency indicator input.41
- Model hub: Model card, PyTorch checkpoint (Hugging Face) (external site: huggingface.co)
- Repository: GitHub repository (external site: github.com)
- Paper: A decoder-only foundation model for time-series forecasting (arXiv 2310.10688) (external site: arxiv.org)
Availability and license
Model-disclosure tier
The model parameters for this release can be downloaded by the public. License terms may still restrict use, redistribution, or commercial use.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Public | PyTorch and Flax checkpoints are published on Hugging Face without an access gate.234 |
| Inference codeIs code for running the model published? | Public | The Apache-2.0 timesfm package provides inference code; the model card includes a PyTorch example.145 |
| Training codeIs the code used to train the model published? | Unknown | A LoRA fine-tuning example (Transformers with PEFT) was added in April 2026; no pre-training code for version 2.5 was found.4 |
| Training-data informationPublic = the training data itself can be obtained. Partial = composition or sources are documented without full access. | Partial | The model card lists GiftEvalPretrain, Wikimedia pageviews (cutoff November 2023), Google Trends top queries (cutoff end of 2022), and synthetic and augmented data. The synthetic data is not documented as released.1 |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Unknown | The ICML 2024 paper describes the original TimesFM approach; the changes in 2.5 are summarized only as a list in the README.14 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Unknown | No evaluation results or evaluation code specific to 2.5 were found in the sources read. |
What it is useful for
The repository describes zero-shot point and quantile forecasting; covariate support through XReg was added for 2.5 in October 2025.4
Run and use notes
- The model card shows loading the checkpoint with the timesfm package (TimesFM_2p5_200M_torch) and compiling a forecast configuration with a maximum context and horizon before forecasting.1
Organization context
Provenance and derivatives
Trained by Google Research as the third open TimesFM checkpoint. The card lists training data from GiftEvalPretrain (Salesforce), Wikimedia pageviews, Google Trends, and synthetic data. No other base model is named.1
Other releases in the TimesFM family
- TimesFM 3.0Model-disclosure tier (USASI rubric v0.1): Open-weight
U.S. eligibility
Eligible · basis: Documented U.S. control
The model card lists Google Research as author. Google Research is part of Google; Alphabet Inc.'s fiscal 2025 Form 10-K lists its principal executive offices in Mountain View, California, and Exhibit 21.01 lists Google LLC as a Delaware subsidiary of Alphabet.167
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.