Tasks and inputs
Granite TimeSeries PatchTST-FM-r2 is listed for Data & control, Time Series Forecasting. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 18, 2026
Granite TimeSeries PatchTST-FM-r2 is IBM's open ~385M zero-shot forecasting model, dual-licensed Apache-2.0/OpenMDW, with 8,192-step context.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights; no per-token price is published. Your cost is the GPU time you run it on.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Forecasting demand for a thousand SKUs rarely justifies training a thousand models. Zero-shot foundation models promise a shortcut: point one pretrained network at any series and get quantile forecasts back. Granite TimeSeries PatchTST-FM-r2 is IBM's latest attempt, and IBM claims it leads permissively licensed zero-shot models on GIFT-Eval as of 31 August 2026.
The result has about 385M parameters, a hidden size of 1024, a training context of 8,192 steps and a quantile head spanning 99 quantiles. Code in IBM's TSFM repository still loads r1 checkpoints.
IBM reports second place on CRPS and MASE among replicable zero-shot models, and third and fourth once models pretrained on GIFT-Eval's training split are included, ahead of some larger systems. The r2 submission was still a pending pull request to the leaderboard when the card was written. Check the live board before quoting a rank.
Dual: OpenMDW 1.0 or Apache-2.0, at your choice. That is unusually clean for commercial forecasting.
Take two years of your own history, hold out the last quarter, and forecast it zero-shot with r2 and with whatever you run now. Compare pinball loss at the 10th, 50th and 90th quantiles, not just point error. If no single model wins every series, IBM's Granite Timeseries Ensemble R1 pools r2 with r1, FlowState and TTM forecasts using equal weights.
IBM's language models are on its provider page; other forecasters are grouped under time-series forecasting.
Compare the model on the work you need to complete.
Granite TimeSeries PatchTST-FM-r2 is listed for Data & control, Time Series Forecasting. It accepts text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Documented limits: 8.19K context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under OpenMDW-1.0 or Apache-2.0 (dual). Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
allenai
ACE2.5-ERA5 is Ai2's open stochastic climate emulator trained on ERA5, producing calibrated atmospheric ensembles from days to centuries at low compute.
Mistral AI
Timestral 1.0 is Mistral's open time-series forecasting family in four sizes (5M-209M), producing zero-shot probabilistic forecasts on CPU.
allenai
HiRO-ACEv2 is Ai2's open AI climate emulator that produces 3 km precipitation, winds and pressure from a 100 km run, over 300x cheaper than a numerical model.
TimesFM 3.0 is a numerical time-series forecasting checkpoint for PyTorch under a custom noncommercial licence.