IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly license

IBM Research released Granite Time Series PatchTST-FM-r2, a foundation model for zero-shot forecasting across demand, prices, energy, traffic, telemetry, and similar series. At roughly 385 million parameters, it supports context up to 8,192 steps, flexible horizons, missing-value imputation, and probabilistic forecasts via a 99-quantile head.

As of Sept. 8, 2026, on GIFT-Eval it ranks second among replicable zero-shot models on CRPS and MASE (geometric-mean CRPS 0.467, MASE 0.6846), immediately behind TimesFM-3—and first in that category among models with commercial-friendly open licenses. Weights, architecture, inference pipeline, and reproduction code are public; dual licensing is Apache 2.0 or OpenMDW 1.0.

For teams tired of one custom model per dataset, this is a practical open alternative that does not force a restrictive license on production forecasts.

We're cheering the boring virtues here. Time-series foundation models only matter if you can deploy them without a legal surprise and without leaking benchmark train sets into the marketing story. IBM documents the pretraining mix—selected GiftEvalPretrain sets, KernelSynth-style synthetics, a TSMixup corpus kept outside GIFT-Eval evaluation data, and about 500,000 CauKer sequences—and still lands near the top even when pretrained (non-strictly-zero-shot) rivals are included: third on CRPS, fourth on MASE among replicable models, beating several larger pretrained systems.

Architecturally, r2 swaps plain transformer blocks for conformer-style layers that pair self-attention with temporal convolution, adds 50% overlapping patches with Hamming-window overlap-add, and grows to 30 blocks. That is a credible explanation for smoother inter-patch forecasts, not just a parameter bump. Early Confluent Cloud early-access hooks for streaming inference via Flink show IBM wants this in production pipes, not only notebooks.

Advocate where earned: permissive SOTA-adjacent forecasting reduces the need for every retailer and grid operator to retrain from scratch. Name the downside: leaderboard crowns move weekly, geometric means hide domain failure modes, and "zero-shot" still needs governance review before it touches inventory or energy trading. Opaque proprietary forecasters will keep winning where liability and data residency demand a vendor throat to choke—unless open weights plus documented corpora keep closing that trust gap.

Our trajectory read: forecasting becomes another foundation-model category where license and leakage hygiene matter as much as a few CRPS points. PatchTST-FM-r2 is a strong open move; treat the GIFT-Eval snapshot as of Sept. 8 as a checkpoint, not a forever ranking.

PatchTST-FM-r2 succeeds PatchTST-FM-r1 in IBM's Granite TSFM family. GIFT-Eval is a broad forecasting benchmark; IBM emphasizes comparisons restricted to replicable zero-shot models without test leakage.