Ant International has released its proprietary Falcon TST (Time-Series Transformer) AI model, the industry-first Mixture of Experts architecture-based big data model with multiple patch tokenizers, supported by up to 2.5 billion parameters.
The company noted that Falcon TST has already been deployed internally at Ant International to manage cashflow and FX exposure on an hourly, daily and weekly basis, and has achieved accuracy rates of over 90%, while cutting the company’s FX costs by up to 60%.
The model has also achieved state-of-the-art zero-shot results on well-acknowledged long-term forecasting benchmarks, such as absolute mean error rates.
Jiang-Ming Yang, Chief Innovation Officer, Ant International, highlighted that Ant International is aiming to “advance the field through global collaboration” by open sourcing the model, which will allow scientists and developers worldwide to contribute real-world feedback through utility.
The model enables businesses to make various other time-series forecasts, such as weather patterns, calendar events, financial market fluctuations, cross-border traffic data, and more.
Ant International has also collaborated with various industry partners to deploy the model in different use cases, such as in helping businesses mitigate FX cost and volatility, as well as to support airlines in offering more stable, competitive pricing to their customers. With the Airports Council International World’s report estimating that global air travel is expected to reach almost 10 billion passengers in 2025, these applications of AI have the potential to translate to significant cost savings for consumers globally.
The Falcon TST AI Model is now available on GitHub and Hugging Face. Ant International invites interested developers to find out more about the model and recent success stories here: https://falcon-tst.ant-intl.com
