What does IBM's new 385M-parameter time-series model change for forecasting?

Zero-shot forecasting that ranks near the top

A foundation model for time series works differently from the models most engineers deploy today. Instead of fitting a separate network to each dataset, you load one pretrained model and ask it to predict the next hours or days of a series it has never seen. IBM's announcement says PatchTST-FM-r2 does exactly that.

On the GIFT-Eval benchmark the model sits second among replicable zero-shot entries as of 8 September 2026. Its geometric-mean CRPS is 0.467 and its geometric-mean MASE is 0.6846. Only TimesFM-3 scores better on CRPS in that restricted leaderboard. When pretrained models that may have seen the evaluation data are added to the comparison, PatchTST-FM-r2 falls to third on CRPS and fourth on MASE — still ahead of Chronos-2, Timer-S1, and several Toto variants, some of which are larger.

The benchmark spans demand, prices, energy loads, traffic, telemetry and other regularly sampled series. Lower values are better for both metrics. The announcement does not say how many individual datasets compose GIFT-Eval, nor does it report wall-clock inference latency.

Conformer blocks replace plain transformers

PatchTST-FM-r1 used standard transformer layers: multi-head self-attention followed by a feed-forward network. The new version swaps those for conformer blocks borrowed from speech processing. Each block wraps the attention layer with two half-step feed-forward layers and inserts a temporal convolution layer.

The convolution kernels alternate in a repeating pattern of 5, 5, 3, 3 across the 30 blocks. That pattern gives the model two complementary tools. Self-attention captures long-range relationships between patches. Convolution provides an inductive bias toward local temporal structure, handling short-distance interactions so attention can focus on longer horizons.

Attention maps on the ETTh1 dataset illustrate the shift. The transformer version concentrates near the diagonal — local relationships. The conformer version shows clear off-diagonal mass, meaning the model attends to distant patches while convolution covers the near ones.

Two more changes smooth the output. Patches now overlap by 50 percent with Hamming-window weighting, and forecasts are stitched together with overlap-and-add. The block count grew from 20 to 30. Together these modifications yield the 385M-parameter count and the 8,192-step context window.

Training corpus spelled out in four parts

Transparency about pretraining data is rare. IBM lists four sources:

• Selected datasets from GiftEvalPretrain
• Custom synthetic data generated with KernelSynth using modified periodic kernels and limited augmentation
• A TSMixup corpus built with the Chronos approach but restricted to datasets outside the GIFT-Eval evaluation set
• Approximately 500,000 synthetic CauKer sequences, each 4,096 steps long

This disclosure is framed as useful for enterprise governance reviews. It does not claim the corpus is free of all licensing risk, nor does it publish the full list of GiftEvalPretrain subsets used.

Probabilistic output and missing-value imputation

The prediction head emits 99 quantiles, not just a point forecast. That lets callers construct prediction intervals at any confidence level. The same pipeline also handles imputation of missing values — a feature the previous model lacked.

Quantile levels are configurable at inference time. The example code requests the 0.1, 0.5 and 0.9 quantiles. The announcement does not specify whether the 99 quantiles are evenly spaced or follow a different grid.

Running it takes three lines of setup

Installation uses the public package:

pip install "granite-tsfm>=0.3.9"

Loading the weights:

model = PatchTSTFMForPrediction.from_pretrained("ibm-granite/granite-timeseries-patchtst-fm-r2")

The forecasting pipeline wraps the model, accepts a pandas DataFrame with a timestamp column and one or more target columns, and returns a forecast DataFrame that includes the requested quantiles. No fine-tuning step is required. The example uses 512 steps of context and a 64-step horizon on the public ETTh1 dataset.

The same weights are also available through an Early Access program on Confluent Cloud, where they run inside Apache Flink for streaming inference. The initial portfolio there includes PatchTST-FM-r1 and FlowState-r1.1. It also includes TTM-r3 and TSPulse. The announcement does not state when the Early Access program opens to general availability.

Backward compatibility and licensing choice

Checkpoints from PatchTST-FM-r1 load in the new architecture implementation. The Granite-TSFM repository hosts the code. Users may pick either Apache 2.0 or OpenMDW 1.0 — both are permissive, both allow commercial use and modification. They also permit redistribution. OpenMDW is a Linux Foundation framework designed specifically for AI artifacts.

No clause-by-clause comparison of the two licenses is provided. It also does not mention whether model cards, datasheets or SBOMs are published alongside the weights.