IBM released Granite Time Series PatchTST-FM-r2, a 385-million-parameter foundation model dual-licensed under Apache 2.0 and OpenMDW 1.0, Hugging Face reported on Wednesday. The release updates the earlier PatchTST-FM-r1 system by introducing a modified architecture, an expanded pretraining corpus, missing-value imputation, and probabilistic forecasting.
Benchmark testing on GIFT-Eval placed PatchTST-FM-r2 second overall among replicable zero-shot models as of September 8, 2026. The model achieved a geometric-mean Continuous Ranked Probability Score of 0.467, trailing only TimesFM-3, and a Mean Absolute Scaled Error of 0.6846. Across models with permissive, commercial licenses, it ranked first. When compared against models allowed to train on benchmark evaluation datasets, it placed third in CRPS and fourth in MASE, outperforming Chronos-2, Timer-S1, and Toto variants.
Conformer blocks adapted from speech processing replace standard transformer layers inside the model. Each block combines two half-step feed-forward layers with multi-head self-attention and a temporal convolution layer. Convolutions use alternating kernel sizes of 3 and 5 in a repeating {5, 5, 3, 3} sequence to capture local temporal patterns, allowing attention heads to track long-range patterns. The network expands from 20 to 30 blocks, incorporates normalization, and processes contexts up to 8,192 steps using 50 percent overlapping patches. A 99-quantile prediction head generates distribution forecasts and uncertainty bounds.
Researchers assembled the pretraining data from four documented sources rather than an uninspected corpus. The data incorporates selected collections from GiftEvalPretrain, synthetic data generated via KernelSynth with modified periodic kernels, a TSMixup corpus excluding GIFT-Eval test sets, and approximately 500,000 synthetic CauKer sequences of 4,096 steps each.
Confluent and IBM integrated earlier Granite time-series models into Apache Flink inside Confluent Cloud through an early-access program. The streaming setup runs inference for PatchTST-FM-r1, FlowState-r1.1, TTM-r3, and TSPulse directly on real-time feeds, avoiding the need to transfer live data to a separate machine-learning cluster.
