.jpg)
Foundational models have transformed the way we approach complex tasks in vision and language, but their application to time series analysis remains a fertile and little-explored field, especially in classification, where the historical focus has been on forecasting. In this context, Mantis emerges, a lightweight model based on transformer architecture that is trained exclusively with synthetic data using self-supervised contrastive learning. Its innovative proposal lies in a tokenization mechanism designed to maximize the potential of transformers, combined with an improved test-time methodology that uses intermediate layer representations, self-ensembling, and embedding fusion. Thanks to these techniques, Mantis achieves superior performance to other foundational models across multiple time series classification dataset collections, demonstrating that it is possible to obtain state-of-the-art results with a lightweight and efficient approach.
Mantis's ability to efficiently process temporal sequences opens new opportunities in sectors where accurate classification of temporal patterns is critical: from anomaly detection in IoT sensors to identifying behaviors in financial series or biomedical signals. In this scenario, having custom applications and platforms that integrate models like this becomes essential. At Q2BSTUDIO we develop custom software and artificial intelligence solutions for companies that allow these innovations to be fully leveraged. Our AWS and Azure cloud services facilitate the scalable deployment of classification models in production, while our capabilities in AI agents and business intelligence services with Power BI help visualize and act on the results obtained.
For a successful implementation of models like Mantis in business environments, having a robust cloud infrastructure is fundamental. From our Azure and AWS cloud services we offer optimized environments for training and inference of deep learning models, ensuring high availability and security. Furthermore, the integration of these models with cybersecurity systems allows detecting anomalous patterns in real time, protecting the organization's critical data.
The evolution toward lightweight foundational models like Mantis represents a paradigm shift: it is no longer necessary to have huge amounts of labeled data to achieve competitive performance. This democratizes access to advanced time series classification techniques. At Q2BSTUDIO, we combine this vision with AI solutions for companies that adapt to each client's specific needs, from implementing AI agents to automating processes based on machine learning. Our focus on custom applications ensures that each solution fits perfectly into the application domain, whether financial, industrial, or healthcare.
Research in time series classification with transformers is taking firm steps toward lighter, more accurate, and more generalizable models. Mantis is a clear example of how the combination of intelligent tokenization, synthetic data, and ensembling techniques can outperform specialized approaches. For companies looking to incorporate these capabilities, having a technology partner that understands both theory and practice is key. At Q2BSTUDIO we offer comprehensive development, cloud, business intelligence, and cybersecurity services, helping organizations transform temporal data into strategic decisions.
