Time-series forecasting has long been a messy, per-dataset chore. Every business line, every sensor type, every region needs its own trained model. Foundation models change that, but two questions decide whether anyone can actually use them in production: does the model generalize zero-shot to data it has never seen, and does the license allow commercial deployment? IBM's Granite Time Series PatchTST-FM-r2, released on September 9, 2026, addresses both questions at once.

The model itself: 385M parameters, conformer blocks

PatchTST-FM-r2 keeps the patch-based representation that made the PatchTST family efficient, but swaps standard Transformer layers for conformer blocks. Inside each block, multi-head self-attention sits between two half-step feed-forward layers, with a temporal convolution layer added in. Self-attention captures long-range dependencies across patches; convolution handles local structure within patches. The two responsibilities no longer fight each other. The backbone grows from 20 to 30 layers; patches now use 50% overlap with Hamming-window weighting, paired with overlap-and-add forecasting to smooth patch boundaries. The final model has roughly 385 million parameters, supports context lengths up to 8192 steps, and a prediction head that emits 99 quantiles, giving both point forecasts and uncertainty intervals.

Where it sits on GIFT-Eval

GIFT-Eval is currently the most comprehensive zero-shot time-series benchmark. As of September 8, 2026, PatchTST-FM-r2 lands at CRPS geometric mean 0.467 and MASE geometric mean 0.6846 in the replicable, strictly zero-shot category — second place overall, behind only TimesFM-3. Since TimesFM-3 ships under a non-commercial license, PatchTST-FM-r2 is effectively the top model in the permissive, commercial-friendly license class. Even when pretrained models that were allowed to include GIFT-Eval training data are added to the comparison, PatchTST-FM-r2 still ranks third on CRPS and fourth on MASE, beating Chronos-2, Timer-S1, and Toto variants while staying a fraction of their size.

Training data: four auditable sources

IBM breaks down the pretraining corpus explicitly. Four sources: a subset of GiftEvalPretrain; synthetic data built on KernelSynth with modified periodic kernels; a TSMixup corpus generated using the Chronos recipe, with all GIFT-Eval evaluation sets excluded; and roughly 500,000 synthetic CauKer sequences of length 4096. The full list is public, not a hand-wave.

Licensing and ecosystem: the lowest possible bar

PatchTST-FM-r2 is dual-licensed under Apache 2.0 and OpenMDW 1.0 — enterprises can pick either. Code, weights, and the inference pipeline all ship via the granite-tsfm repository on GitHub and Hugging Face, and the inference code stays backward-compatible with PatchTST-FM-r1 checkpoints. On the streaming side, IBM has already wired PatchTST-FM-r1, FlowState-r1.1, TTM-r3, and TSPulse into Flink on Confluent Cloud under an Early Access program, putting the foundation models directly into stream processing pipelines.

The short version

PatchTST-FM-r2 is not just another leaderboard climb. It drops the commercial-use bar to Apache 2.0, ships the architectural changes as open code rather than a paper claim, and borrows conformer blocks from speech models in a way that translates cleanly to time series. For teams that need real production forecasting without getting blocked by licensing, this is one of the few options that genuinely works today.