IBM's 385M-Param PatchTST-FM-r2 Ranks 2nd on GIFT-Eval Zero-Shot

IBM released Granite Time Series PatchTST-FM-r2 on September 9, 2026. As of September 8, 2026, it is the highest-performing zero-shot model under a permissive commercial license on the GIFT-Eval leaderboard, ranking 2nd overall among replicable, strictly zero-shot models by both CRPS and MASE. When pretrained models — those allowed to include GIFT-Eval evaluation dataset splits in their corpora — are added to the comparison, PatchTST-FM-r2 places 3rd for CRPS and 4th for MASE, outperforming Chronos-2, Timer-S1, and Toto variants despite competing without that benchmark training data. Weights, architecture code, inference pipeline, and benchmark-reproduction scripts are all public.

Architecture: Conformer Blocks Replace Standard Transformer Layers

PatchTST-FM-r2 keeps the patch-based input representation from its predecessor but replaces standard transformer blocks with conformer-style blocks. Each conformer block arranges two half-step feed-forward layers around a multi-head self-attention sublayer and a temporal convolution layer — a structure with roots in speech processing. Self-attention handles long-range inter-patch dependencies; convolution handles local temporal structure, freeing attention from near-diagonal concentration. Attention-pattern comparisons on ETTh1 samples confirm this: transformer blocks show heavy near-diagonal concentration, while conformer blocks exhibit pronounced long-distance, off-diagonal focus. The backbone uses alternating convolution kernel sizes of 3 and 5 in a repeating pattern of {5, 5, 3, 3}.

Additional changes relative to r1: 50% overlapping patches with Hamming-window weighting, overlap-and-add forecasting to smooth patch boundaries, added normalization for stability, and an expansion from 20 to 30 blocks. The result is approximately 385 million parameters, a maximum context length of 8,192 steps, flexible forecast lengths, and a 99-quantile prediction head delivering both point forecasts and full distributional outputs. The architecture is backward-compatible with PatchTST-FM-r1 checkpoints via the Granite-TSFM repository.

GIFT-Eval Benchmark Results

Lower values are better for both CRPS and MASE. Chronos-2, Timer-S1, and Toto are pretrained models permitted to include GIFT-Eval training splits in their corpora; PatchTST-FM-r2 beats all three in the combined leaderboard without that data.

Training Corpus and Licensing

The pretraining corpus draws from four documented sources: selected datasets from GiftEvalPretrain; custom synthetic data built on KernelSynth with modified periodic kernels and limited augmentation; a TSMixup corpus generated using the Chronos methodology but restricted to datasets outside the GIFT-Eval evaluation set; and approximately 500,000 synthetic CauKer sequences each of length 4,096. No GIFT-Eval evaluation splits appear in training, maintaining clean zero-shot status.

The model is dual-licensed under Apache 2.0 and OpenMDW 1.0 — the Linux Foundation's framework designed for AI models and related materials — with users free to choose either. Documented corpus provenance gives enterprises conducting model-governance and IP review considerably more to work with than an opaque pretraining description would provide.

Deployment: Python and Confluent Streaming

Getting started requires pip install "granite-tsfm>=0.3.9", after which the model loads from ibm-granite/granite-timeseries-patchtst-fm-r2 on the Hugging Face Hub via PatchTSTFMForPrediction.from_pretrained. The TimeSeriesForecastingPipeline wrapper accepts a context length (512 in the documented example), a prediction length (64 in the example), quantile levels, and a frequency string, then generates forecasts with no fine-tuning. Requested quantiles — 0.1, 0.5, and 0.9 in the sample code — are returned alongside point forecasts.

For streaming workloads, IBM and Confluent have made several Granite Time Series models available through an Early Access program on Confluent Cloud, integrating inference into Apache Flink pipelines so forecasts and anomaly-detection results can be generated from live data streams without a separate batch ML environment. The current Early Access portfolio includes PatchTST-FM-r1, FlowState-r1.1, TTM-r3, and TSPulse.

As architectural specificity continues to drive state-of-the-art results, PatchTST-FM-r2's combination of a documented corpus, dual permissive license, reproducible benchmarks, and a streaming integration path removes several adoption blockers that have historically pushed teams toward closed alternatives.