PatchTST-FM R1
Model directoryibm-research/granite-timeseries-patchtst-fm-r1~325M params | Apache-2.0 | Upstream checkpoint
PatchTST-FM R1 is IBM's foundation-model extension of PatchTST, trained for broad zero-shot forecasting rather than a single dataset. Where the original upstream PatchTST checkpoint is a narrow ETTh1 baseline, PatchTST-FM R1 turns the patch-based architecture into a general-purpose forecaster, and IBM positions it as a top-5 replicable zero-shot model on the GIFT-Eval leaderboard as of March 2026.
At roughly 325M parameters it pairs an 8192-token context window with a 99-quantile probabilistic output head, and it can perform imputation and forecasting simultaneously, so series with missing values are handled in a single pass rather than requiring a separate fill step. It is pretrained on a broad multi-domain corpus aimed at zero-shot generalization; IBM positions the model on the GIFT-Eval leaderboard without disclosing the full training mixture.
Where to run this model
Compute availability, hardware requirements, and pricing depend on the execution provider.
Get model weights, deployment instructions, and current provider listings.
- Retrocast · The Forecasting Company
managed service · Family / provider variant · Source checked 2026-10-06
Forecast through the managed API using patchtst-fm. The provider controls the deployed variant and request limits.
- Confluent Cloud · Early Access
managed service · Family / provider variant · Source checked 2026-10-06
IBM lists PatchTST-FM-r1 in the Confluent integration. Early access is for evaluation, not production.
Hugging Face provider mappings checked 2026-10-06. No Inference Provider mapping was listed for this checkpoint.
Model Classification
Family
PatchTST-FM / Granite TimeSeries
Type
time series foundation model
Pretrained time-series model for zero-shot or few-shot forecasting workloads.
Resources
Training Data
Broad multi-domain pretraining corpus for zero-shot generalization; IBM positions the model on the GIFT-Eval leaderboard without disclosing the full mixture.
Recommended For
- • Zero-shot probabilistic forecasting across diverse domains
- • Long-context forecasting with simultaneous imputation of missing values
Strengths
- • 8192-token context enables very long lookback windows
- • 99-quantile output provides rich uncertainty estimates without sampling
Limitations
- • Large model footprint requires GPU inference
- • Newer foundation-model release with less production track record than the original PatchTST baseline
Capabilities
Tags
Specifications
- Parameters
- ~325M
- Architecture
- PatchTST transformer with 8192-token context and 99-quantile output head
- Context length
- n/a
- Max context
- n/a
- Minimum history
- n/a
- Recommended history
- n/a
- Input step
- n/a
- Required target series
- 1
- Temperature
- Ignored
- Top P
- Ignored
- Max output
- n/a
- License
- Apache-2.0