TL;DR
Play games included with Prime
Start a Prime free trial and play with Amazon Luna on your devices.
Start playingAs an affiliate, we earn on qualifying purchases.
IBM has released Granite Time Series PatchTST-FM-r2, a roughly 385 million-parameter model built for zero-shot forecasting, missing-value imputation and probabilistic predictions. IBM reports that it ranked highest among permissively licensed, replicable zero-shot models on GIFT-Eval as of September 8, 2026, although real-world performance will depend on each deployment.
IBM has released Granite Time Series PatchTST-FM-r2, a roughly 385 million-parameter forecasting model offered under Apache 2.0 or OpenMDW 1.0. IBM says the model led the permissively licensed entries in a defined category of the GIFT-Eval benchmark as of September 8, giving businesses and developers a new option for zero-shot time-series forecasting without task-specific training.
PatchTST-FM-r2 is designed to forecast data such as demand, prices, energy loads, traffic and telemetry. It supports input histories of up to 8,192 time steps, flexible forecast lengths, missing-value imputation and probabilistic output through a 99-quantile prediction head. Users can obtain point forecasts as well as ranges representing uncertainty.
On GIFT-Eval, IBM reports a geometric-mean CRPS of 0.467 and a geometric-mean MASE of 0.6846. Lower values are better for both measures. The model ranked second on both metrics when IBM limited the comparison to replicable zero-shot systems evaluated without test leakage, and it ranked highest within that group among models carrying what IBM describes as permissive commercial licensing. Those positions apply to the leaderboard and comparison conditions reported on September 8, 2026.
IBM has published the model weights, architecture and inference pipeline, along with code intended to reproduce the benchmark results. The model is dual-licensed under Apache 2.0 and OpenMDW 1.0, allowing users to select either license. The architecture implementation is available through IBM’s Granite-TSFM repository and is described as backward-compatible with PatchTST-FM-r1 checkpoints.
A new zero-shot forecasting contender
IBM’s Granite Time Series PatchTST-FM-r2 pairs a roughly 385 million-parameter forecasting backbone with probabilistic predictions, missing-value imputation, open weights and a choice of commercial-friendly licenses.
One model, three forecasting jobs
The release targets general-purpose time-series work across demand, pricing, energy load, traffic and telemetry—without requiring task-specific training before the first prediction.
Zero-shot predictions
Generate forecasts for previously unseen series without fitting a separate model to every target dataset. Flexible horizons support varied planning windows.
Missing-value imputation
Fill gaps in incomplete sequences before or alongside downstream analysis—a practical feature for operational data with interrupted collection.
Probabilistic output
A 99-quantile prediction head returns both point forecasts and uncertainty ranges for decisions that depend on plausible best- and worst-case outcomes.
R2 reworks the forecasting backbone
The patch-based family remains, but standard transformer layers give way to conformer-style blocks that combine local pattern detection with longer-range attention.
Local detail meets long context
Temporal convolution focuses on nearby signals while multi-head self-attention models relationships across longer spans. The architecture remains backward-compatible with PatchTST-FM-r1 checkpoints.
Strong reported scores—with boundaries
Lower scores are better for both reported measures. Rankings apply only to IBM’s stated comparison conditions and the leaderboard snapshot dated September 8, 2026.
| Evaluation claim | Reported result | Interpretation | Evidence status |
|---|---|---|---|
| Geometric-mean CRPS | 0.467 | Probabilistic forecast quality | ~ Vendor reported |
| Geometric-mean MASE | 0.6846 | Scaled point-forecast error | ~ Vendor reported |
| Replicable zero-shot group | ✓ Second overall | No test leakage condition | ✓ Reproduction code released |
| Permissively licensed subset | ✓ Highest ranked | Defined commercial-friendly group | ~ Scope-limited claim |
| Independent production audit | ✗ Not identified | Private enterprise workloads | ✗ Still untested |
Benchmark standing does not establish production reliability, inference cost, latency, memory use or business value. Each deployment requires its own validation.
Open artifacts improve testability—not certainty
IBM has published model weights, architecture, inference tooling and benchmark-reproduction materials through Hugging Face and the Granite-TSFM repository.
The credible upside
- Open weights and documented implementation make the system inspectable.
- Apache 2.0 or OpenMDW 1.0 broadens potential commercial reuse.
- Zero-shot use may reduce the need for many dataset-specific models.
- Quantile forecasts can support inventory, capacity and energy decisions under uncertainty.
The production gap
- No comparative figures are reported for latency, memory demand or operating cost.
- Private datasets may contain irregular sampling and regime shifts absent from benchmarks.
- No peer-reviewed evaluation or independent production audit is identified.
- Organizations remain responsible for legal review, governance and consequential-use validation.
“Top performing zero-shot model released under a permissive, commercial-friendly open-source license.”
IBM Research · Hugging Face announcementFrom leaderboard to live data
The decisive evidence will come from outside teams reproducing the results and testing calibration, accuracy and cost on unseen commercial workloads.
Verify the scores
Run IBM’s published pipeline and confirm the stated GIFT-Eval results without test leakage.
Use private data
Compare zero-shot accuracy with specialist baselines on genuinely unseen operational series.
Price the runtime
Record latency, throughput, hardware needs, memory use and total serving cost.
Test calibration
Confirm that prediction intervals remain reliable during volatility and distribution shifts.
A useful release, with deployment claims still to earn
The combination of competitive reported results, open artifacts and permissive licensing makes PatchTST-FM-r2 easier to inspect and test than an opaque forecasting service. That is a meaningful advantage—but independent replication and production evidence should decide whether it becomes the right model for a specific business.
Permissive Licensing Broadens Deployment
The combination of competitive zero-shot results and broad reuse rights could make PatchTST-FM-r2 relevant to organizations that want a general forecasting model but cannot accept restrictive terms. Zero-shot operation may reduce the work needed to build separate models for every dataset, although teams will still need to test the system against their own operational data.
The release also provides probabilistic forecasts, which can be more useful than a single predicted value when decisions depend on uncertainty. Inventory planning, capacity management and energy operations often need a range of plausible outcomes. Still, a benchmark rank does not establish production reliability, cost or business value, and those questions remain deployment-specific.
As an affiliate, we earn on qualifying purchases.
R2 Reworks the Forecasting Backbone
PatchTST-FM-r2 follows IBM’s earlier PatchTST-FM-r1 while retaining the family’s patch-based representation of time-series data. The new version replaces standard transformer layers with conformer-style blocks that combine multi-head self-attention and temporal convolution. IBM says this design lets convolution handle nearby patterns while attention focuses on longer-range relationships.
The revision expands the network from 20 to 30 blocks and introduces overlapping patches, Hamming-window weighting and overlap-and-add forecasting. IBM also documented four groups of pretraining data, including selected GiftEvalPretrain datasets, modified KernelSynth data, a restricted TSMixup corpus and about 500,000 synthetic CauKer sequences, each with a length of 4,096 steps.
“top performing zero-shot model released under a permissive, commercial-friendly open-source license”
— IBM Research, writing on Hugging Face
As an affiliate, we earn on qualifying purchases.
Production Performance Still Untested
It is not yet clear how closely the GIFT-Eval rankings will match results across private enterprise datasets, irregular sampling patterns or rapidly changing conditions. The announcement does not supply comparative figures for inference speed, memory use, operating cost or fine-tuned performance, all of which may affect model selection.
The benchmark discussion comes from IBM Research’s own announcement, although the company has released artifacts intended to support replication. No peer-reviewed evaluation or independent production audit is identified in the announcement. Organizations also retain responsibility for license review, data governance and validation before deploying the model in consequential workflows.
As an affiliate, we earn on qualifying purchases.
Independent Tests Move to Real Data
Developers can now download PatchTST-FM-r2 from Hugging Face and run it through IBM’s Granite TSFM package. The immediate test will be whether outside teams can reproduce the published scores and obtain comparable accuracy on data that differs from the benchmark corpus.
Attention will also turn to latency, hardware requirements and forecast calibration in production. IBM and Confluent already have an early-access effort covering several other Granite time-series models for streaming applications, but the announcement does not say when or whether PatchTST-FM-r2 will join that offering.
As an affiliate, we earn on qualifying purchases.
Where I land
My assessment is that this is a credible and useful release, mainly because IBM paired competitive reported results with open weights, documented training inputs and permissive licenses. That package makes the model easier to inspect and test than a forecasting service whose training data, implementation or usage rights remain opaque. This is interpretation, not a reported finding.
The strongest counterargument is that benchmark standing can overstate practical value. A model may perform well across an aggregate evaluation yet be slower, more expensive or less accurate than a smaller specialist model on one company’s data. The vendor-authored presentation and lack of reported independent production evidence keep me from treating the ranking as a broad verdict on deployment quality.
I would raise my assessment if outside evaluators reproduced the scores across several hardware setups and published strong results on unseen commercial datasets, including latency and calibration. I would lower it if replication exposed data contamination, unstable forecasts or costs that erased the benefits of zero-shot use.
Source: Hugging Face
Key Questions
What is Granite Time Series PatchTST-FM-r2?
It is an IBM time-series foundation model with roughly 385 million parameters. It is built to produce forecasts for new time series without task-specific training and can also return quantile-based uncertainty estimates.
Does IBM claim the model leads every forecasting benchmark?
No. IBM’s ranking claim is limited to permissively licensed, replicable zero-shot models on GIFT-Eval as of September 8, 2026. It ranked second overall within the broader replicable zero-shot comparison reported by IBM.
Can businesses use the model commercially?
IBM offers the model under Apache 2.0 or OpenMDW 1.0, which it describes as permissive and suitable for commercial use. Businesses should still conduct their own legal and governance review for a planned deployment.
Does the model require fine-tuning?
The model is designed for zero-shot forecasting, so users can generate predictions without fitting it to each target dataset. Fine-tuning may still be explored, but the announcement centers on performance without task-specific training.
Are the benchmark results independently verified?
The reported rankings come from IBM Research’s Hugging Face publication. IBM has released code and other artifacts for reproduction, but the announcement does not identify a peer-reviewed study or independent production audit.
Source: Hugging Face
NFL season / tailgating Picks
team gear
As an affiliate, we earn on qualifying purchases.
