The Pulse
IBM Releases 385M-Parameter Granite Forecasting Model
IBM has released Granite Time Series PatchTST-FM-r2, a 385-million-parameter model for zero-shot forecasting, probabilistic prediction and missing-value imputation. The model ranks second among replicable zero-shot systems on GIFT-Eval and
AI.info Team ·
IBM’s new model targets forecasts without task-specific training
IBM has released Granite Time Series PatchTST-FM-r2, a new foundation model designed to forecast unfamiliar time series without fine-tuning on each individual dataset. The roughly 385-million-parameter system supports point and probabilistic forecasts, missing-value imputation and flexible prediction horizons.
The release appeared on September 9, 2026, in a post on Hugging Face authored by Roman Vaculin, Wesley M. Gifford, Jiri Navratil, Chandra Reddy and Ayhan Sebin of IBM Research. The model weights, inference pipeline, architecture implementation and code for reproducing the benchmark results are available through the Granite Time Series model page.
“Instead of training and maintaining a separate model for every dataset, users can use a pretrained model and generate forecasts zero-shot,” Roman Vaculin, principal research staff member at IBM Research, wrote in the release.
IBM positions the model for workloads including demand planning, energy consumption, traffic, transaction volumes, prices, computer-system telemetry and industrial sensors. Users provide a recent history of observations, and the model produces forecasts without a task-specific fitting step.
PatchTST-FM-r2 ranks behind TimesFM-3 on GIFT-Eval
IBM’s headline performance claim comes from GIFT-Eval, a benchmark covering time-series forecasting across varied datasets and prediction settings. As of September 8, the model ranks second among replicable, zero-shot systems evaluated without test leakage.
PatchTST-FM-r2 records a geometric-mean continuous ranked probability score, or CRPS, of 0.467 and a geometric-mean mean absolute scaled error, or MASE, of 0.6846. Lower values are better for both measures. IBM says the model sits immediately behind TimesFM-3 in the zero-shot CRPS comparison and ranks first among systems released under permissive commercial licenses.
The benchmark position comes with limits that IBM identifies in its own documentation. GIFT-Eval includes models classified as pretrained, allowing them to use training portions of evaluation datasets in their pretraining data. When those systems are added to the comparison, PatchTST-FM-r2 ranks third for CRPS and fourth for MASE among replicable models.
The model card also says the results were submitted through a pending pull request to GIFT-Eval as of August 31. The Hugging Face announcement uses the September 8 leaderboard state for its comparison, so the ranking describes a dated benchmark snapshot rather than a permanent standing.
Conformer blocks replace the standard transformer stack
PatchTST-FM-r2 keeps the patch-based approach used by the earlier PatchTST-FM-r1 but changes the model’s internal blocks. Standard transformer layers are replaced with conformer-style blocks that combine multi-head self-attention with temporal convolution.
The arrangement places a convolution layer between two half-step feed-forward layers. Attention can model relationships across distant patches, while convolution handles shorter-range patterns in adjacent portions of a series. IBM says the design is intended to reduce the tendency of patch-based forecasts to produce discontinuities between neighboring patches.
The model uses overlapping patches of length 16 with a stride of eight. Training applies Hamming-window weighting, while inference uses an overlap-and-add method to combine forecasts from neighboring patches. IBM also increases the architecture from 20 blocks in r1 to 30 blocks in r2 and adds a pre-head normalization layer for training stability.
PatchTST-FM-r2 accepts a context length of up to 8,192 observations and uses a 99-quantile prediction head. That head allows the system to return forecast distributions rather than only a single expected value, giving users several estimates across different probability levels.
IBM lists four sources for the pretraining corpus
The model’s documented pretraining data combines several sources. IBM says it used selected datasets from GiftEvalPretrain, custom synthetic sequences based on KernelSynth, a TSMixup corpus generated with a process described in the Chronos work, and approximately 500,000 synthetic CauKer sequences with a length of 4,096.
IBM says the TSMixup data was restricted to datasets outside the GIFT-Eval evaluation set. That separation is intended to limit test leakage, one of the conditions used in the benchmark comparison. The model card lists the expanded synthetic data as one of the main differences from PatchTST-FM-r1.
Publishing the data sources gives researchers more information for assessing the model’s benchmark claims and suitability for their own applications. It does not remove the need to check whether a particular deployment meets an organization’s governance, data-use and licensing requirements.
Apache 2.0 and OpenMDW 1.0 make the weights broadly usable
IBM is releasing PatchTST-FM-r2 under a dual-license arrangement. Users can choose either the Apache License 2.0 or the Open Model Development and Distribution License version 1.0, commonly called OpenMDW 1.0.
The permissive licensing is central to IBM’s positioning of the model. The company says PatchTST-FM-r2 is the highest-performing model in its GIFT-Eval category among systems with commercial-friendly licenses. The architecture is also available in IBM’s public Granite TSFM repository, which IBM says remains backward compatible with PatchTST-FM-r1 checkpoints.
Developers can install the package with pip install "granite-tsfm>=0.3.9", load the model from the Hugging Face Hub and pass a recent data window to the forecasting pipeline. IBM’s example uses 512 hourly observations from the ETTh1 dataset to produce a 64-step forecast with 10th, 50th and 90th percentile outputs.
The release extends IBM’s wider Granite Time Series portfolio. IBM and Confluent announced on September 1 that several Granite time-series models were entering Early Access in Confluent Cloud, where forecasts and anomaly results can be generated from streaming data through Apache Flink. PatchTST-FM-r2 is released first as an open model for direct experimentation, with its benchmark scores, implementation and licensing terms available for inspection.