Language modeling had two defining moments. First, a single pretrained model, fine-tuned per task, displaced task-specific systems. Then scale made fine-tuning optional for many tasks, letting the same model handle them from few or no examples. Behavioral models learn from sequences of human actions, such as purchases, payments, and app sessions, but remain organized around individual datasets, tasks, and domains. We present BehaviorGPT-v4, a 12.5B-parameter Large Behavioral Model pretrained by next-event prediction on 150 billion user actions (3.6 trillion tokens) from retail, engagement, and payments. It represents behavior as sequences of discrete, multimodal events and extends next-token prediction to an open set of candidates, still scored in a single step.
On public datasets whose catalogs and users were held out from pretraining, BehaviorGPT-v4 with no target-specific training outperforms every sequential, generative, transferable, tabular, and LLM baseline we evaluated, including specialists trained on the target’s full training split. Fine-tuned, it learns from added target data 17 times faster than the strongest baseline. The same model serves three industries and 13 datasets with no per-dataset changes: it leads the strongest baseline on 30 of 31 dataset–task pairs without adaptation and on all 31 fine-tuned. The 12.5B model transfers better zero-shot than our 0.5B model and needs less target data. It ranks a full catalog in one forward pass, in under a millisecond per query, 48 times faster than a prompted LLM. Production A/B tests of the BehaviorGPT family show double-digit gains in conversion and orders in several deployments. We document the recipe and design decisions behind the BehaviorGPT family, with internal ablations.
An SDK and API for zero-shot deployment and evaluation on new catalogs, and live demos built on them, are available at www.unboxai.com/behaviorgpt.
@techreport{unbox2026behaviorgptv4,
author = {Rickard {Br{\"u}el-Gabrielsson} and Vasudev Gupta and Tom Boustedt and Adam Fredriksson and Simon Granstr{\"o}m and Marcel R{\o}d and Gon{\c{c}}alo Marques and Jens Palmborg and Erik Guander and Nicolas Sanchez and Artem Lukoianov and Srijan Sood and Georgios Kolovos and Simon Ejdemyr and Toni Rosino and Mattias Holmstr{\"o}m and Marc Mathieu and Pratik Thakar and Wen Yao and Miguel Paredes and Peter Sarlin and Alexander Statnikov and John {Br{\"u}el-Gabrielsson} and Steve Flinter and Bryan McCann and Sebastian Siemiatkowski and Gunnar Carlsson},
title = {BehaviorGPT-v4: One Large Behavioral Model for Retail, Engagement, and Payments with Zero-Shot Transferability},
institution = {Unbox AI},
year = {2026},
month = oct,
url = {https://research.unboxai.com/behaviorgpt-v4-one-large-behavioral-model.html},
}