As of September 8, the model was the top permissively licensed replicable zero-shot entry on the GIFT-Eval benchmark and ranked second overall among replicable zero-shot models, with a geometric-mean CRPS of 0.467 and MASE of 0.6846, sitting just behind TimesFM-3. IBM emphasises the ranking is a snapshot rather than a permanent claim on the lead.
Conformer architecture with quantile forecasts
The model moves from standard Transformer blocks to conformer blocks, combining attention with temporal convolution, and uses 50 percent overlapping Hamming-weighted patches. It stacks 30 blocks with context up to 8,192 steps and ends in a 99-quantile prediction head, giving probabilistic forecasts rather than single point estimates.
The pretraining mix draws on selected GiftEvalPretrain datasets, KernelSynth-style synthetics, a TSMixup corpus restricted to data outside the GIFT-Eval evaluation set, and about 500,000 CauKer sequences of length 4,096. IBM has published the weights, architecture and a reproduction pipeline on the Hugging Face Hub under the name ibm-granite/granite-timeseries-patchtst-fm-r2.
Why zero-shot forecasting matters
Zero-shot forecasters can make predictions on datasets they were never trained on, reducing the need for domain-specific model training. For businesses dealing with demand forecasting, energy load prediction or financial time series, a strong open model could cut both costs and development time.
The release continues IBM's Granite strategy of open, enterprise-friendly models with transparent training data and reproduction pipelines. With permissive licensing, the model is positioned for commercial use as well as academic research.
