NVIDIA Transaction Foundation Model

Card-transaction embedding demo

Gradio + ZeroGPU Space for NVIDIA’s ~29M Llama decoder pretrained on TabFormer sequences — domain tokenization, 512-d last-token embeddings, and behavioral similarity scoring.

This Space is staged as static until a ZeroGPU community grant is approved. The Gradio app (app.py) is already in the repo and ready to run once hardware is attached.

• NVIDIA blog
• Hub model mirror
• Upstream GitHub