TimeMixer++
Model directorythuml/timemixer-pp~50M params | Upstream checkpoint
TimeMixer++ is THUML's multi-scale, multi-resolution time-series pattern machine and the successor to the ICLR 2024 TimeMixer architecture. Rather than specializing in a single objective, it is built as a general analytical model that supports 8 tasks including forecasting, classification, anomaly detection, and imputation.
Architecturally it is a multi-scale multi-resolution MLP-Mixer. It employs Multi-Resolution Time Imaging (MRTI) and Time Image Decomposition (TID) to extract patterns across both the time and frequency domains, then combines them through unified multi-scale mixing (MCM) and multi-resolution mixing (MRM) strategies that capture both fine-grained and macro-level structure. It is trained on multi-domain corpora spanning ETT, Weather, ECL, Traffic, Solar, PEMS, and the M4 competition datasets, totaling 10B+ observations.
Where to run this model
Compute availability, hardware requirements, and pricing depend on the execution provider.
Get model weights, deployment instructions, and current provider listings.
The public Hub listing could not be verified (401) on 2026-10-06. Consult the upstream research links for current availability.
Model Classification
Family
TimeMixer
Type
time series foundation model
Pretrained time-series model for zero-shot or few-shot forecasting workloads.
Resources
Training Data
Multi-domain time-series corpora spanning ETT, Weather, ECL, Traffic, Solar, PEMS, and M4 competition datasets — 10B+ observations.
Recommended For
- • Multi-task time-series analysis where you need forecasting, classification, and anomaly detection from one model
- • Multi-resolution workloads spanning different temporal granularities
Strengths
- • Broadest multi-task coverage among MLP-Mixer-based models in the catalog
- • Multi-scale mixing captures both fine-grained and macro-level patterns
Limitations
- • MLP-Mixer backbone may plateau on very long contexts compared to full-attention Transformers
- • Less established in production deployments than the largest Transformer-based families
Capabilities
Tags
Specifications
- Parameters
- ~50M
- Architecture
- multi-scale multi-resolution MLP-Mixer (MCM + MRM)
- Context length
- n/a
- Max context
- n/a
- Minimum history
- n/a
- Recommended history
- n/a
- Input step
- n/a
- Required target series
- 1
- Temperature
- Ignored
- Top P
- Ignored
- Max output
- n/a
- License
- n/a