TTM: Tiny Foundation Models for Multivariate Time-Series Forecasting

2026-06-22 · Source: Deep Learning on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, long

Summary

Tiny Time Mixers (TTMs) are a new family of compact pre-trained models designed for zero-shot and few-shot multivariate time-series forecasting. Built on the lightweight TSMixer architecture, TTMs incorporate adaptive patching, diverse resolution sampling, and resolution prefix tuning to handle heterogeneous multi-resolution pre-training. The models employ a multi-level strategy, pre-training the backbone channel-independently and fine-tuning a decoder for cross-channel correlations and exogenous effects. TTM variants, starting from around 1M parameters (TTM-Base), achieve strong transfer learning performance. For instance, TTMA (5M parameters) outperforms Moirai by 4–10% and TimesFM by 19% in zero-shot forecasting. TTMB (1M parameters) outperforms Chronos by 17–32% and Lag-Llama by 40%. Pre-trained on 1B samples using 6 A100 GPUs for 24–30 hours, TTMB demonstrates significantly faster inference, requiring only 0.01 seconds per batch on CPU, making it suitable for resource-constrained environments.

Key takeaway

For MLOps Engineers deploying time-series forecasting solutions, TTMs offer a compelling alternative to large foundation models. You can achieve strong zero-shot or few-shot performance with significantly reduced computational overhead, enabling faster inference on CPU-only or resource-constrained systems. Consider TTM-Base for its 1M parameters and 0.01 seconds/batch CPU inference. This approach allows you to adapt pre-trained models efficiently to specific multivariate datasets, including those with exogenous variables, without expensive full-model fine-tuning.

Key insights

Tiny Time Mixers (TTMs) enable efficient, accurate time-series forecasting via compact pre-trained models and multi-level adaptation.

Principles

• Channel-independent pre-training generalizes temporal dynamics.
• Resolution diversity is critical for time-series foundation models.
• Lightweight decoders adapt to target-specific multivariate structures.

Method

TTMs pre-train a TSMixer-based backbone on diverse, multi-resolution data using adaptive patching, diverse resolution sampling, and resolution prefix tuning, then fine-tune a small decoder for target-specific multivariate adaptation.

In practice

• Use TTMs for zero-shot or few-shot time-series forecasting.
• Employ the exogenous mixer for known future covariates.
• Adapt pre-trained models to new forecast lengths via pruning.

Topics

• Multivariate Time Series
• Foundation Models
• Zero-shot Forecasting
• Few-shot Learning
• TSMixer Architecture
• Model Efficiency

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Zero-shot Multivariate Time Series Forecasting Using Tabular Prior Fitted Networks — Multivariate time series forecasting can be reframed as scalar regression for zero-shot tabular foundation model application.
• Time-Series Foundation Model Embeddings for Remaining Useful Life Estimation — Frozen time-series foundation models like Chronos-2 enable data-efficient RUL estimation with a lightweight regression head.
• MSMixer: Learned Multi-Scale Temporal Mixing with Complementary Linear Shortcut for Long-Term Time Series Forecasting — Multi-scale MLP architectures with dynamic weighting and linear shortcuts improve long-term time series forecasting.
• The digital quant: instant portfolio optimization with JointFM — JointFM is a zero-shot foundation model predicting full multivariate distributions for instant, domain-agnostic time-series forecasting.
• Five Questions About Chronos-2, the Time Series Foundation Model — Time series foundation models like Chronos-2 enable zero-shot forecasting by learning recurring temporal shapes from diverse data.
• GITCO: Gated Inference-Time Context Optimization in TSFMs — Patch-based TSFMs can be improved at inference time by optimizing input context to suppress harmful patches.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Deep Learning on Medium.