NuFormer
#3 in Cloud-KI & MLOpsNubank · 2× · tolest 24. Aug. 2026
68
Momentum
NuForcer is a proprietary, internally developed AI Foundation Model from Brazilian digital bank Nubank, based on Transformer architecture and self-supervised learning. It was developed to learn generalizable representations from decades of transaction and behavioral data for use in risk modeling, credit lending, fraud detection, and product recommendations. The first generation comprised 330 million parameters and was trained on approximately 600 billion Tokens; the model is not a separate commercial SaaS product, but runs internally in Nubank's credit decision and personalization systems in Brazil, Mexico, and Colombia.
Momentum-Verloop
26.05.24.08.
Features
| Platform | Operated internally at Nubank on proprietary GPU compute infrastructure; latest generation architecture similar to open-source models Kimi K3 and Qwen3.5 |
| Price | No public pricing – internal proprietary system, not sold as a standalone product |
| Compute Performance (FLOPS/TOPS) | No official FLOPS/TOPS figures available; newest generation reportedly quadrupled context length, training and inference speed |
| Release Date | First generation developed over 12-15 months, unveiled in Q3 2025 earnings call (Nov. 2025) |
| Memory | Trained on approx. 600 billion tokens (first generation, 330M-parameter model); full dataset spans trillions of tokens |
| Availability | In production for credit card underwriting in Brazil and Mexico, and unsecured lending in Brazil; expansion to Colombia planned |
