IBM Granite Time Series PatchTST-FM-r2 Release

IBM has released Granite Time Series PatchTST-FM-r2, a general-purpose foundation model for zero-shot time-series forecasting. This model enables users to generate forecasts for demand, prices, energy loads, and telemetry without the need to train separate models for every dataset.

State-of-the-Art Zero-Shot Performance

PatchTST-FM-r2 is the highest-performing zero-shot model released under a permissive, commercial-friendly license among replicable models on the GIFT-Eval leaderboard as of September 8, 2026. It ranks #2 overall among replicable, zero-shot models for both Continuous Ranked Probability Score (CRPS) and Mean Absolute Scaled Error (MASE).

Even when compared against "pretrained" models—those permitted to include GIFT-Eval training data in their pretraining corpora—PatchTST-FM-r2 remains highly competitive, ranking 3rd for CRPS and 4th for MASE. It outperforms several larger pretrained models, including Chronos-2, Timer-S1, and Toto variants.

Architectural Enhancements

PatchTST-FM-r2 evolves the architecture of its predecessor, PatchTST-FM-r1, to better capture both long- and short-term temporal relationships.

Conformer-Based Backbone

The model replaces standard transformer layers with conformer blocks. These blocks combine multi-head self-attention with a temporal convolution layer, providing two complementary reasoning mechanisms:

  • Self-attention: Models long-range relationships between patches.
  • Convolution: Provides an inductive bias toward local temporal structure, capturing shorter-term interactions.

To optimize this, the conformer blocks use alternating convolution kernel sizes of 3 and 5 in a repeating {5, 5, 3, 3} pattern. This allows the self-attention mechanism to focus on long-distance relationships rather than concentrating near the diagonal, as is common in standard transformers.

Refined Patching and Stability

To improve forecasting accuracy and smooth patch boundaries, the model implements:

  • 50% overlapping patches with Hamming-window weighting.
  • Overlap-and-add forecasting.
  • Increased depth: The architecture has expanded from 20 to 30 blocks.
  • Normalization: Added for improved stability.

Model Specifications and Capabilities

PatchTST-FM-r2 is a ~385M-parameter model with the following technical capabilities:

  • Context Length: Supports up to 8,192 steps.
  • Probabilistic Forecasting: Predicts 99 quantiles via a specialized prediction head, providing both point forecasts and uncertainty intervals.
  • Flexible Forecast Lengths: Supports variable lengths for future predictions.
  • Missing Value Support: Includes support for the imputation of missing values.

Training Data and Transparency

IBM provides a documented pretraining corpus to assist enterprise users with governance and licensing reviews. The training data consists of four primary sources:

  1. Selected datasets from GiftEvalPretrain.
  2. Custom synthetic data based on KernelSynth with modified periodic kernels and limited augmentation.
  3. A TSMixup corpus (restricted to datasets outside the GIFT-Eval evaluation set).
  4. Approximately 500,000 synthetic CauKer sequences of length 4,096.

Licensing and Availability

Granite Time Series PatchTST-FM-r2 is dual-licensed under Apache 2.0 and OpenMDW 1.0, allowing users to choose the license that best fits their needs. Both licenses provide permissive rights to use, modify, and distribute the model.

Model weights, architecture, the inference pipeline, and the code required to reproduce benchmark results are available via the Hugging Face Hub and the granite-tsfm GitHub repository. The implementation is backward-compatible with PatchTST-FM-r1 checkpoints.

Production Integration

Beyond standalone use, IBM and Confluent have integrated Granite Time Series models into an Early Access program in Confluent Cloud. This allows foundation-model inference to be performed directly within streaming applications using Apache Flink, enabling real-time forecasting and anomaly detection on live data streams without moving data to a separate ML environment.

Sources