Post-training case study: Reaching frontier-level intelligence on global payroll with Deel

AI21 Labs and Deel collaborated to post-train the NVIDIA Nemotron 3.5 Super VL model for the specialized domain of global payroll. Using a small dataset, they employed a two-step training process involving near-policy Supervised Fine-Tuning (SFT) and Self-Distillation Policy Optimization (SDPO). This methodology significantly improved the model's ability to provide complete and accurate answers to complex payroll questions. The resulting model surpassed the performance of the base Nemotron model by over 20 points and outperformed the more expensive Claude Opus 5.5 model by 8 points, at a 50x lower cost.