Time-series data comes in many forms, and with many potential applications. That means no single forecasting method can work best all the time.

If you’re trying to predict tomorrow’s high and low temperature, or whether a company will hit next week’s sales target, point forecasting is a good bet. But if you’re trying to decide when to restock a product or evaluate a company’s risk exposure, a probabilistic forecast could be more useful. Other times, you may be trying to detect anomalies in a stream of real-time data, to prevent a network disruption or machinery break down, and speed is essential.

IBM Research has built a family of time-series foundation models that shine in each scenario. Released this week, the models are currently at the top of the GIFT-Eval leaderboard in Hugging Face: FlowState-r1.1, for point forecasting (among zero-shot, replicable, models without data leakage), PatchTST-FM-r1, for probabilistic forecasting (among replicable, non-agentic models without data leakage), and TTM-r3 and TSPulse-r1, for efficient forecasting, anomaly detection, classification and search, supporting thousands of inferences per second.

The models are built on distinct architectures, each with their own strengths. In this blog post, we break down what’s new, what kinds of tasks each model is ideally suited for, and how they can be accessed and used.

Our newest transformer based model, PatchTST-FM-r1, co-developed with Rensselaer Polytechnic Institute, evolved from our ground-breaking PatchTST architecture, which introduced channel independence and patching for more accurate and efficient forecasting. PatchTST-FM-r1 model excels at forecasting tasks. It handles context lengths ranging from 128 to 8,192 timepoints, enables forecasting over both short and long time horizons, and is robust to missing values. Its current GIFT-Eval rank demonstrates that transformer architectures can be flexible, expressive, and perform with high accuracy when trained on large-scale datasets consisting of real and synthetic data.

TTM-r3 is the third generation of our TinyTimeMixer models, designed to balance speed and accuracy. This release introduces several innovations that improve forecasting accuracy and speed up inferencing by 15 to 50 times over today’s state-of-the-art models. TTM-r3 also thrives in CPU-only environments, making it well-suited for real-world, high-throughput deployments. TTM-r3 supports rapid fine-tuning, forecasting with many variables, and can even incorporate control variables, improving its performance in complex, industrial scenarios. Our TTM models are extremely popular, with more than 37 million downloads on Hugging Face so far.

FlowState-r1.1 is the newest version of our FlowState model, which is built on a novel state-space architecture called S5, which can handle short and long inputs and forecasting horizons, as well as varied sampling rates. By combining a state space model encoder with a functional-basis decoder, FlowState-r1.1 has the rare ability to harmonize data with varying sampling rates to produce accurate long-horizon forecasts. This new version incorporates additional synthetic training data and increases context length to further improve performance.

All three model releases expand IBM’s time-series model portfolio, and complement TSPulse, our family of compact pre-trained models that excel at anomaly detection, search, classification, and imputation tasks.

Designed for enterprise deployments

Together, these models cover a wide range of real-world, enterprise requirements. Applications include industrial manufacturing and monitoring, as well as detecting IT incidents and electrical grid disruptions. Collectively, the models have been trained on more than 100 billion data points taken from public domains or synthetically generated.

The table below highlights each model’s key strengths by application. For high throughput and low latency on a CPU machine, try TTM-r3. For accurate point and probabilistic forecasting, try FlowState-r1.1 and PatchTST-FM-r1, respectively. And for time series anomaly detection, classification and other non-forecasting tasks, try TSPulse.

How we got here

IBM has led the research and development of time-series foundation models since the release of TST in 2021, which was among the first transformer-based models applied to time-series data. In 2023, PatchTST introduced the concepts of patching and channel independence to time series data, making it efficient and effective for transformers to process long time series. In 2023, TSMixer blazed the way for greater speed and efficiency with the use of multi-level perceptrons (MLPs) to combine patch and cross-channel information.

Building on the mixers, Tiny Time Mixer (TTM), released in 2024, introduced the first lightweight time series models for tasks across many domains. In 2025, building on state-space models (SSMs), an efficient type of recurrent neural network, we introduced the FlowState time-series model. Its innovations included parallel training, functional basis decoding, and sampling-rate invariance. Earlier versions of TTM and FlowState were top performers on GIFT-Eval for forecasting in 2024 and 2025.

This year, we complete our portfolio with the release of PatchTST-FM-r1, TTM-r3, and FlowState-r1.1, which once again are top performers on GIFT-Eval at forecasting. They are complemented by TSPulse, which specializes in anomaly detection and classification. Together, these models cover a wide spectrum of real-world use cases.

Try out IBM’s time-series models

All our models are open weight and can be downloaded in Hugging Face.

• Our research versions are available under a non-commercial license. • Our family of Granite time-series models have been trained on curated datasets and are available under a permissive Apache 2.0 license.

Several notebooks are available to help users get started with IBM’s time-series models. The notebooks highlight each model’s capabilities and best use-cases, and are built on top of model architecture and supporting code from our open-source repository.