IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly license
2026-09-09 · Hugging Face
IBM Releases SOTA Granite Time Series PatchTST-FM-r2 Model
On September 9, 2026, IBM released the Granite Time Series PatchTST-FM-r2 model, the latest addition to the Granite TSFM family. Time-series foundation models are transforming forecasting systems by allowing users to generate forecasts zero-shot using a pretrained model, eliminating the need to train and maintain separate models for every dataset.
Key Model Features
PatchTST-FM-r2 combines an updated architecture, a larger pretraining corpus, and strong zero-shot performance within a ~385M-parameter model. Its main characteristics include:
- Parameters: Approximately 385 million parameters.
- Context Length: Supports a context length of up to 8,192.
- Forecasting: Offers flexible forecast lengths and probabilistic forecasting through a 99-quantile prediction head.
- Applications: General-purpose zero-shot forecasting for demand, prices, energy loads, traffic, telemetry, and other time series.
- Imputation: Supports the imputation of missing values.
Architecture: Changes from PatchTST-FM-r1
PatchTST-FM-r2 retains the patch-based representation of its predecessor but redesigns the internal architecture to efficiently capture long- and short-term relationships and smooth inter-patch predictions.
- Conformer Layers: The model transitions from standard transformer layers to conformer layers, which incorporate convolution along with multi-head self-attention. This architecture originated in speech processing.
- Dual Reasoning Mechanisms: A PatchTST-FM-r1 block combined multi-head self-attention with a feed-forward network. In r2, this was replaced with a conformer-style block containing two half-step feed-forward layers surrounding multi-head self-attention and a temporal convolution layer. This provides complementary reasoning: self-attention models long-range relationships between patches, while convolution provides an inductive bias toward local temporal structure to capture shorter-term interactions.
GIFT-Eval Benchmark Performance
A foundation model is most valuable when it generalizes to untrained time series. GIFT-Eval provides a broad evaluation across heterogeneous datasets.
- Zero-Shot Performance: As of September 8, 2026, among replicable, zero-shot models evaluated without test leakage, PatchTST-FM-r2 ranks second for both CRPS and MASE. It achieved a geometric-mean CRPS of 0.467 and MASE of 0.6846.
- Top Commercial-Friendly Model: Importantly, PatchTST-FM-r2 is the highest-performing model in this category among those with permissive, commercial-friendly open-source licensing.
- Competitive Against Pretrained Models: Even when compared to "pretrained" models allowed to include GIFT-Eval training data in their corpora, PatchTST-FM-r2 remains near the top (3rd for CRPS and 4th for MASE). It outperforms models like Chronos-2, Timer-S1, and Toto variants, despite some competitors being considerably larger.
Open Source and Licensing
The model is dual-licensed under Apache-2.0 and OpenMDW-1.0, allowing users to select either permissive, commercial-friendly license.
Model weights, architecture, the inference pipeline, and code to reproduce benchmark results are all openly available. Users can access and try the model on Hugging Face. Furthermore, models from the Granite Time Series family can be leveraged in streaming applications in production settings using Confluent products.