Wednesday, September 9, 2026
Tech Beat
Sep 9, 2026, 3:36 PMArtificial Intelligence

IBM Granite PatchTST-FM-r2 Leads Commercially Licensed Zero-Shot Forecasting Models

IBM's 385M-parameter Granite PatchTST-FM-r2 ranks first among permissively licensed zero-shot models, adding probabilistic forecasts and data imputation.

Listen to this briefingAudio briefing

Summary

IBM released Granite Time Series PatchTST-FM-r2 on September 9, 2026, offering its weights, architecture, inference pipeline and reproducibility code through Hugging Face and Granite-TSFM. As of September 8, it ranked second among replicable, leakage-free zero-shot models on GIFT-Eval for CRPS and MASE, behind TimesFM-3, but first among models with permissive commercial licensing. Its geometric-mean CRPS was 0.467 and MASE was 0.6846. Against replicable models allowed to train on benchmark data, it placed third for CRPS and fourth for MASE, outperforming Chronos-2, Timer-S1 and Toto variants.

The roughly 385 million-parameter model handles contexts up to 8,192 steps, flexible forecast lengths, missing-value imputation and point or probabilistic forecasts through a 99-quantile head. Compared with PatchTST-FM-r1, IBM replaced transformer layers with conformer blocks combining self-attention, temporal convolution and two half-step feed-forward layers, expanded from 20 to 30 blocks, and added normalization, alternating convolution kernels of 3 and 5, plus 50% overlapping Hamming-weighted patches. Pretraining combines four documented sources, including selected GiftEvalPretrain data, modified KernelSynth data, GIFT-Eval-excluded TSMixup data and about 500,000 synthetic CauKer sequences of length 4,096. Users can choose Apache 2.0 or OpenMDW 1.0. Separately, Confluent Cloud Early Access supports PatchTST-FM-r1, FlowState-r1.1, TTM-r3 and TSPulse for live forecasting and anomaly detection through Apache Flink, but does not yet list r2.

Positives

  • PatchTST-FM-r2 ranks first among permissively licensed, replicable zero-shot models on GIFT-Eval and second overall for both CRPS and MASE.
  • A 0.467 geometric-mean CRPS and 0.6846 MASE place IBM immediately behind TimesFM-3 in leakage-free zero-shot testing.
  • The 99-quantile head supplies point forecasts, probability distributions and uncertainty intervals without task-specific fitting or fine-tuning.
  • Apache 2.0 and OpenMDW 1.0 licensing permits broad commercial use, modification and distribution.
  • Open weights, architecture, inference code and benchmark reproduction materials improve technical and governance transparency.
  • Contexts up to 8,192 steps, missing-value imputation and flexible forecast lengths broaden applications across demand, energy, traffic, prices and telemetry.

Risks & concerns

  • TimesFM-3 still leads PatchTST-FM-r2 in the replicable, leakage-free zero-shot GIFT-Eval comparison.
  • Including models allowed to use benchmark training data drops IBM to third for CRPS and fourth for MASE.
  • IBM says documented training data does not replace customers’ own model-governance and licensing reviews.
  • Confluent Cloud’s Early Access portfolio lists PatchTST-FM-r1, FlowState-r1.1, TTM-r3 and TSPulse, not the new r2 model.
Primary sourceHugging Face - Bloghttps://huggingface.co/blog/ibm-research/ibm-releases-sota-granite-time-series
Read full article
Editorial note: Tech Beat summarizes and analyzes third-party reporting. The source link is the authoritative article. This page does not reproduce the full source text.

More From The Wire

Artificial IntelligenceSep 9

Paul Christiano Joins OpenAI Foundation Board and Safety Committee

Artificial IntelligenceSep 9

Anthropic Researcher Jacob Coxon Quits, Warns Self-Improving AI Could Kill Humanity

Artificial IntelligenceSep 9

ControlAI Urges Superintelligence Halt After OpenAI Hugging Face Breach