IBM ships PatchTST-FM-r2, the top zero-shot time-series forecaster under a permissive license
On September 9, 2026, IBM released Granite Time Series PatchTST-FM-r2, a 385M-parameter model that becomes the best zero-shot forecaster shipped under a permissive license on the GIFT-Eval benchmark, ahead of several larger models. For any team doing demand, load or telemetry forecasting, it is a production-ready zero-shot starting point.
September 9, 2026. IBM released Granite Time Series PatchTST-FM-r2, the latest iteration of its time-series foundation-model family. September 8, 2026. On the GIFT-Eval leaderboard, the model became the best zero-shot forecaster released under a permissive license — Apache 2.0 and OpenMDW 1.0 — and ranks 2nd overall among replicable zero-shot models. September 6, 2026. Forecasting outgrows the “one model per dataset” era. Why it matters: forecasting demand, load or telemetry is leaving the bespoke age, and the license no longer blocks deployment.
A foundation model that forecasts without ever seeing your data
A time-series foundation model works like a large language model: instead of training and maintaining a separate model for every series — sales, electricity consumption, network traffic — you deploy a pretrained model that generates zero-shot forecasts, on data it has never seen. PatchTST-FM-r2 applies that logic to demand, prices, energy loads, traffic and telemetry.
The model weighs in at roughly 385 million parameters, accepts a context length up to 8,192 points, produces flexible forecast horizons and, crucially, emits probabilistic forecasts through a 99-quantile prediction head. That last detail matters: a point forecast is not enough in operations — you need a confidence interval to act on it. Native support for missing-value imputation rounds out the picture, since real series have gaps and a model that handles them natively removes a preprocessing step.
The architecture: conformer blocks rather than a plain transformer
The break with PatchTST-FM-r1 is not in parameter count — it is in the backbone. The model keeps the patch-based representation that made the PatchTST family effective — the series is cut into segments treated as tokens — but its internal skeleton is rebuilt around conformer blocks.
A conformer combines multi-head self-attention with a temporal convolution. Self-attention captures long-range dependencies — seasonal trends, cycles — while the convolution captures short-range structure between neighboring points. It is that dual read, IBM says, that “smoothens inter-patch predictions” and substantially improves the error metrics. In plain terms: the model reads both the overall shape of the series and its local grain, without sacrificing either.
From a single point to an interval — and from r1 to r2
A point forecast — “tomorrow’s demand will be 4,200” — says nothing about the risk of being wrong. PatchTST-FM-r2’s 99-quantile prediction head changes the nature of the output: for each horizon the model emits a distribution — the median value and the extreme quantiles that bound the uncertainty — rather than a bare number. A capacity-planning team reads the P90 quantile to size with a deliberate margin; a procurement team reads the P10 to negotiate without overstocking. It is the difference between answering “how much?” and “how much, and with what confidence?”
The leap from PatchTST-FM-r1 comes down to three changes rather than a simple parameter bump: a backbone rebuilt around conformer blocks (self-attention for long-range dependencies, convolution for local grain), a larger pretraining corpus, and probabilistic forecasting. Native support for missing-value imputation rounds it out: production series have gaps, and a model that handles them at inference spares an upstream cleaning pipeline.
Scores that beat larger models
On GIFT-Eval, the reference benchmark that evaluates forecasters across heterogeneous datasets, PatchTST-FM-r2 posts a geometric-mean CRPS of 0.467 and a MASE of 0.6846 — lower is better. Restricted to models that are zero-shot, replicable and evaluated without test leakage, it ranks 2nd on both metrics, just behind TimesFM-3.
The number that actually matters is elsewhere: among zero-shot models under a permissive, commercial-friendly license, PatchTST-FM-r2 is first. And it stays competitive even when you add “pretrained” models allowed to include the benchmark’s training data in their corpora: 3rd on CRPS, 4th on MASE, ahead of Chronos-2, Timer-S1 and Toto variants — several of them considerably larger. A 385M-parameter model outpacing bigger rivals is the signature of an efficient architecture rather than raw capacity.
The license changes the game, not just the score
Benchmarks tend to fix attention on rank, but for a team that has to put a model into production, the license is usually the first filter. PatchTST-FM-r2 is dual-licensed — Apache 2.0 and OpenMDW 1.0, the user picks either — both permissive and compatible with royalty-free commercial use.
That is what sets it apart from rivals that are technically excellent but locked behind non-commercial licenses or inaccessible weights. IBM publishes the weights, the architecture, the inference pipeline and the code to reproduce the benchmark results. A team can download the model, serve it on its own infrastructure, and evaluate its fit on their own series without a legal review or a dependency on a third-party API. For forecasting — a cross-cutting need that rarely touches the intellectual property of training data — that openness is a deployment argument in its own right.
From notebooks to streaming
IBM positions PatchTST-FM-r2 beyond the batch use case: the Granite Time Series family integrates with Confluent for streaming forecasting. The idea is to attach the model to live telemetry or measurement streams, producing forecasts on the fly rather than in nightly batches.
That is the direction operational forecasting is moving: an anomaly in electric load, a spike in traffic or a drift in consumption is caught continuously, not the next morning. A zero-shot foundation model makes that real-time shift credible, because it does not require retraining a model per metric — you plug it in, it forecasts, and you re-evaluate it periodically.
A practical caveat worth stating: zero-shot strength on a benchmark is a starting hypothesis, not a guarantee. A foundation model’s real fit shows in the gap between its forecasts and your actuals on a held-out slice of your own history. The disciplined path is to run PatchTST-FM-r2 against your last quarter of data, compare it to a naive seasonal baseline, and retire a bespoke model only once the foundation model beats it on your metric — not on GIFT-Eval’s.
Verdict
PatchTST-FM-r2 is not an absolute record, and it does not claim to be. Its value is a rare combination: a top-tier zero-shot score under a permissive license, with open weights and reproduction code — which makes it genuinely deployable, not merely citable in a paper.
If you carry a forecasting workload — demand planning, load forecasting, infrastructure telemetry — use PatchTST-FM-r2 as your zero-shot baseline before considering a bespoke model. Evaluate it on your own series: SOTA on GIFT-Eval does not guarantee domain fit, and a 99-quantile output requires interpreting prediction intervals correctly, not just the central value.
If your constraint is licensing or data sovereignty, this model checks both boxes — Apache 2.0/OpenMDW 1.0 and local inference — where many competitors impose an API or a non-commercial license. In 2026, that is the difference between a model you watch and a model you ship.