IBM released Granite Time Series PatchTST-FM-r2, a ~385M-parameter zero-shot forecasting foundation model with 8,192-token context, 99-quantile probabilistic forecasts, and conformer-based architecture, published September 9, 2026 on Hugging Face.

PatchTST-FM-r2 scores geometric-mean CRPS 0.467 and MASE 0.6846 on GIFT-Eval — #2 among replicable zero-shot models and #1 under permissive licensing, beating pretrained Chronos-2, Timer-S1, and Toto. Critically, r1 was CC-BY-NC-SA-4.0 (non-commercial); r2 is dual-licensed Apache-2.0 / OpenMDW-1.0.

The license flip from non-commercial CC-BY-NC-SA to Apache-2.0/OpenMDW means production forecasting systems can now ship the top permissive-licensed zero-shot model without legal work — a single ~3 GB self-hostable checkpoint replacing per-dataset model training.

IBM Granite PatchTST-FM-r2: Top Zero-Shot Forecasting Model Under Apache-2.0

What shipped

IBM released Granite Time Series PatchTST-FM-r2 on Hugging Face today: a ~385M-parameter time-series foundation model for zero-shot forecasting (demand, prices, energy loads, telemetry) with probabilistic forecasts via a 99-quantile prediction head, 8,192-token context, and imputation support. I pulled the repo metadata to verify: 384,600,208 F32 parameters, ~3.07 GB, last modified September 9, 2026.

What changed

The headline numbers from the announcement: geometric-mean CRPS of 0.467 and MASE of 0.6846 on GIFT-Eval — #2 among replicable zero-shot models (behind only TimesFM-3), and #1 in that category among permissively licensed models. Even when models allowed to train on benchmark data join the comparison, it stays 3rd (CRPS) and 4th (MASE), beating pretrained Chronos-2, Timer-S1, and Toto variants. Architecture-wise, r1's transformer blocks became conformer blocks — attention for long-range structure, temporal convolution for the local inductive bias, with alternating kernel sizes of 3 and a larger one.

The change nobody's highlighting is the license. I checked both model repos via the HF API: ibm-research/patchtst-fm-r1 is CC-BY-NC-SA-4.0 — non-commercial. r2 is dual-licensed Apache-2.0 / OpenMDW-1.0 (the repo card ships openmdw-1.0). Same lineage, 49% more parameters (257.9M → 384.6M), and the "NC" gone. Weights, pipeline, and benchmark reproduction code are all public.

Why a builder cares

Time-series forecasting is one of the last ML domains where companies still hand-train per-dataset models — it's also a domain where "open weights" often quietly means "research only." A zero-shot model that tops the permissive-license leaderboard at ~385M params (single ~3 GB safetensors file, trivially self-hostable) makes the commercial path a config change instead of a license negotiation. r1 got 39.6k downloads under a license most businesses couldn't ship; r2 already has 2.4k downloads in its first day. If you forecast anything in production, this is now the default open option to benchmark against.

What shipped

What changed

Why a builder cares

Related