IBM launches Granite PatchTST as time-series forecasting shifts to open foundational models

IBM has launched the new time-series model 'Granite PatchTST-FM-r2' (PatchTST-FM-r2), the latest release in the Granite family of temporal forecasting models. The model weighs roughly 385 million parameters and supports a temporal context of up to 8,192 steps, offering advanced probabilistic forecasting and generating estimates based on 99 percentile intervals, as well as automatic missing-value imputation. The release marks a shift in forecasting system architecture, allowing technical teams to produce instant-measurement forecasts without prior training, rather than having to train and maintain a separate model for each data set.

The model ranks first among reproducible, open-commercial-license models on the standard benchmark platform 'GIFT Eval'.According to the leaderboard data up to 8 September 2026, the model placed second overall among reproducible zero-shot models on both error metrics, scoring 0.467 on CRPS and 0.6846 on MASE, directly behind 'Times FM 3', making it the top performer among all open-commercial-license models. Even when compared with models that were allowed to incorporate evaluation data into their pre-training, it retained third place on CRPS and fourth on MASE, outperforming models such as 'Chronos 2', the 'Toto' variants and 'Timer S 1', although some competing models are larger in parameter count.

The improvement stems from a redesign of the internal architecture compared with the previous version, as IBM replaced standard transformer blocks with hybrid blocks inspired by the 'Conformer' architecture widely used in audio signal processing. The new block combines multi-head self-attention layers with temporal convolution layers whose kernels alternate between sizes 3 and 5 in a repeating pattern. This design gives the model two complementary mechanisms: convolution layers capture short-term temporal structure and relationships, while self-attention models long-range dependencies across thousands of time steps. The architecture also incorporates Hamming windows with 50 % overlap between temporal patches to reduce edge prediction errors, increases the number of blocks from 20 to 30, and adds normalization layers to improve training stability.

IBM adopted a transparent training methodology to meet corporate governance requirements, using a publicly disclosed dataset comprised of selected pre-training sources from the GIFT Eval platform, modified synthetic data, and roughly 500 thousand synthetic sequences of 4,096 steps each. The model is released under a dual license, allowing a choice between the Apache 2.0 license and the OpenMDW 1.0 license from the Linux Foundation, both granting broad rights for commercial use, modification, and redistribution without restrictive conditions. IBM also provided a Python library and the full model weights on the Hugging Face platform, alongside cloud integration with the Confluent platform to run real-time forecasts and anomaly detection directly on live data streams via the Apache Flink engine without moving data to separate machine-learning environments.

This offering opens a practical path for data and technology teams in the Gulf and Egypt to reduce the cost of running enterprise forecasting.In critical sectors such as electric load management, cooling plants, and water desalination in the Gulf, or supply-chain and logistics services at regional ports and airports, institutions typically have to maintain hundreds of statistical models that consume extensive time for tuning and retraining. The availability of a lightweight 385 million-parameter model that operates in a silent-forecast mode under a full commercial license enables banks and local firms to run it locally within their national data centers or private servers without risking data leakage or incurring monthly fees for closed-source APIs. Directly coupling the model with streaming data processing moves analytics in distribution networks and fintech platforms from retrospective issue monitoring to proactive prediction of bottlenecks and liquidity before they occur.

The availability of this model demonstrates that the foundation-model revolution goes beyond text and images to reach the core of operational data and sensitive sensors. This trend gives developers and engineers in the Arabic tech ecosystem a precise, open-source tool for re-engineering real-time forecasting and building advanced enterprise applications without relying exclusively on closed foreign platforms.