Model registry
Foundation Models for Time Series Classification
Explore 10 pretrained models evaluated in TSC-FM Benchmark, with classifier configurations, papers, code and checkpoint references. Compare in-context classification with frozen representation learning.
| Model | Organization | Parameters | Evaluated configurations | Paper | Code | Checkpoint |
|---|---|---|---|---|---|---|
| ChorusTIC In-Context Learning | GDUT · Huawei Noah’s Ark Lab · Université Paris Cité · MBZUAI · Shantou University | Not reported by authors | ChorusTIC.ckpt (step 6000) Native in-context inference | Paper | Code | Weights |
| Mantis Representation Learning | Huawei Noah’s Ark Lab · Université Paris Cité | 8.11M backbone | Mantis-8M kNN LR MLP NC RF ScaledLR SVM | Paper | Code | Weights |
| MantisV2 Representation Learning | Huawei Noah’s Ark Lab · Université Paris Cité | 4.19M backbone | MantisV2 LR RF ScaledLR | Paper | Code | Weights |
| MOMENT Representation Learning | Auton Lab, Carnegie Mellon University | 125M backbone | MOMENT-1-base kNN LR MLP NC RF SVM | Paper | Code | Weights |
| NuTime Representation Learning | Microsoft Research · Peking University | 2.4M backbone | checkpoint_bias9.pth RF | Paper | Code | Weights |
| TabICL In-Context Learning | Inria · École Normale Supérieure–PSL | Not reported for v1.1 checkpoint | tabicl-classifier-v1.1-20250506.ckpt Native in-context inference | Paper | Code | Weights |
| TabICLv2 In-Context Learning | Inria · École Normale Supérieure–PSL | Not reported for v2 checkpoint | tabicl-classifier-v2-20260212.ckpt Native in-context inference | Paper | Code | Weights |
| TIC-FM In-Context Learning | GDUT · Huawei Noah’s Ark Lab · Université Paris Cité · MBZUAI | Not reported by authors | TSEncoder_orion_icl_fullv1.0.pt Native in-context inference | Paper | Code | Weights |
| TiCT In-Context Learning | Visa Research | 47M (TiCT-Large, ResNet) | No standalone checkpoint link found Native in-context inference | Paper | Code | No standalone checkpoint link found |
| UniShape Representation Learning | SCUT · A*STAR · Khalifa University | 3.1M backbone | unishape_checkpoint_zeroshot.pth RF | Paper | Code | Weights |
Parameter counts marked as backbone-only exclude dataset-specific classifier heads. “Not reported” is used when an official source does not publish the full count.
How are foundation models used for time series classification?
A time series foundation model (TSFM) learns reusable representations or prediction behavior during pretraining. Time series classification (TSC) assigns a class label to an input sequence. This benchmark compares how pretrained models transfer to classification through two approaches: in-context inference and frozen features with a classifier head.
A univariate sequence has one measured channel; a multivariate sequence has several. Check each model’s input type and the results for the relevant dataset scope before comparing scores. Catalog membership does not imply complete coverage of every evaluation setting.
In-context classification
These configurations predict from labeled context examples without fitting a classifier for each target dataset. The catalog includes time series models as well as TabICL and TabICLv2, tabular foundation models applied to vectorized time series.
Frozen representations and classifier heads
These configurations extract representations with a frozen pretrained encoder and use a downstream classifier. The classifier head is part of the evaluated configuration: two heads on the same backbone can produce different results.
How to compare time series classification models
Use the time series classification leaderboard to compare average accuracy, average rank and dataset coverage under the same label budget and modality scope. The catalog groups classifier heads under one backbone; the leaderboard treats each evaluated configuration separately.
Standard evaluation uses the official training split. For limited-label comparisons, select Low-shot and then 5-shot or 10-shot. These settings use different feasible dataset subsets, so a higher score across settings does not by itself establish a better model. Read the few-shot classification protocol before interpreting the results.
Overall scores summarize performance across datasets. To investigate where models differ, use the capability analysis by dataset attributes, which evaluates its own fixed comparison core. Model detail pages provide configuration results and source references for further inspection.
Choose the evaluation setting that matches your task
- If you need predictions from labeled examples without fitting a target-specific classifier, inspect the in-context learning configurations.
- If you can fit a classifier while keeping the pretrained encoder fixed, inspect frozen-probe configurations. Compare classifier heads as separate configurations, even when they share a backbone.
- If only a few labels are available per class, use the matching 5-shot or 10-shot view and check dataset coverage before drawing conclusions from aggregate accuracy.
The step-by-step evaluation guide explains how to select a dataset scope, label budget and prediction approach before reading the leaderboard.