IBM released Granite Time Series r2 on Hugging Face, and it ranks first among permissively licensed forecasters on GIFT-Eval

IBM Research published Granite Time Series PatchTST-FM-r2 on the Hugging Face blog on 9 September 2026. The post announces the second-generation model in IBM's PatchTST foundation-model family, with weights, architecture, inference pipeline and training code all released.

The model and the licence

IBM says PatchTST-FM-r2 is about 385 million parameters. The architecture uses conformer blocks that combine multi-head self-attention with temporal convolution, with 50 percent overlapping patches weighted by a Hamming window. The prediction head outputs 99 quantile levels, so it produces a full probability distribution for each forecast step, and the context window is up to 8,192 steps. The repository is backward-compatible with r1 checkpoints, so a team already running r1 can swap the weights.

The licence is what makes this release different from most time-series foundation models on Hugging Face. IBM ships the model under a dual choice of Apache 2.0 and OpenMDW 1.0, both permissive, and the announcement highlights that users can pick either one for commercial use.

The benchmark IBM cites

IBM reports that on the GIFT-Eval benchmark, as evaluated on 8 September 2026, PatchTST-FM-r2 ranks second among "zero-shot replicable" models on both metrics the benchmark reports: 0.467 CRPS and 0.6846 MASE, both geometric-mean scores. Among models with permissive licences, IBM claims the top position on both metrics.

When the comparison widens to include pretrained models that were allowed to train on GIFT-Eval data, IBM says PatchTST-FM-r2 slips to third on CRPS and fourth on MASE, and names Chronos-2, Timer-S1, and the Toto family as models it outperforms. Those numbers all come from IBM's own post.

What it is aimed at

The named use cases are demand forecasting, price forecasting, energy load, traffic, sensor telemetry, CPU utilisation and transaction volumes. The model handles missing-value imputation as well as forecasting, and IBM points to a Confluent Cloud Early Access integration for streaming inputs, so a team already piping metrics through Kafka can forecast on the same stream.

IBM says the release includes the weights, the training data mix, the training recipe, and the inference pipeline and code. A team that wants to fine-tune on private data therefore has every part of the chain and does not have to reverse-engineer the model from the checkpoint alone.

For a team already running a paid forecasting API or a heavier foundation model, the practical question is whether a 385 million parameter open model at the top of the permissive-licence table is enough to move the workload off a vendor and onto their own hardware. IBM's post is written to argue that it is, and every number in it comes from IBM, so a team weighing that switch should run its own evaluation on its own data before committing.

Source

• IBM releases SOTA Granite Time Series PatchTST-FM-r2 model with commercial-friendly license (IBM Research on Hugging Face, 9 September 2026)
• Model: ibm-granite/granite-timeseries-patchtst-fm-r2
• Code: ibm-granite/granite-tsfm