Custom foundation models import support for Granite Time Series models

IBM has expanded its Custom Foundation Models feature to support Granite Time Series models (TinyTimeMixer and TTM), enabling practitioners to import their fine-tuned multivariate TTM forecasting models directly into watsonx.ai and use the Timeseries Model Inferencing API/SDK. Granite Time Series models are lightweight, open-source models optimized for forecasting.

Bringing your TTM—tuned for multivariate domain customization—into watsonx.ai unlocks enterprise-grade governance, seamless API integration and scalable deployment workflows, while leveraging the power of your enterprise data.

IBM’s Granite Time Series Tiny Time Mixer (TTMs) are compact models for multivariate time-series forecasting, open-sourced by IBM Research under the Apache 2.0 license.

Pretrained TTMs with 1-5 M parameters were previously made available on watsonx.ai and shown to deliver state-of-the-art zero-shot forecasting accuracy across a variety of datasets, ranging from IoT sensor readings to energy demand and financial time series, while running efficiently even on CPU-only machines. These models support multiple input context lengths (from 512 to 1536 timepoints), making them versatile for a wide range of forecasting scenarios

With the addition of support for custom TTMs, users can now fine-tune on their own data, capturing the correlation between multiple channels as well as support for exogenous features, and then bring these models to the watsonx.ai platform across different industry use cases. 

1. Enterprise-grade governance and security: By importing your tuned TTM into watsonx.ai, you leverage the platform’s built-in governance features such as audit trails, role-based access controls and compliance checks, ensuring your forecasting models meet organizational policies and regulatory requirements.
2. Seamless API/SDK integration: The new Time Series Model Inferencing API/SDK allows you to interact with your custom TTM model programmatically, integrating forecasting into your data pipelines, dashboards or AI agent workflows with just a few lines of code.
3. End-to-end AI developer platform: Within watsonx.ai, you can further enhance your forecasting workflows using hosted LLMs and frameworks to build agents, RAG for context enrichment and one-click deployment, all within a unified environment to build AI solutions integrating timeseries forecasting.
4. Domain customization: Offline finetuning of the TTM allows users to fine-tune the TTM models for their domain specific data accounting for semantics of inputs, controls, and output KPIs and bring this fine-tuned model for improved performance.

Bring your Granite Time Series TTM models into an enterprise-grade AI platform and discover how easy it is to deploy your tuned time-series models at scale.

Start your free watsonx.ai trial today

Explore our sample notebooks