AI21 Post-Trains NVIDIA Nemotron 3.5 Super VL with Gong, Deel, and Alice for Specialized Enterprise Models

AI21’s Forward-Deployed Researchers (FDRs) post-trained the recently released NVIDIA Nemotron 3.5 Super VL in partnership with Gong, Deel, and Alice. The initiative demonstrates how companies can transition from renting general models to owning specialized intelligence using AI21’s Lab-as-a-Service and proprietary Harness platform, completing the process in weeks rather than quarters.

Gong: Matching Frontier Quality at Lower Cost

Gong, a revenue AI company, needed to reduce costs for its Gong Data Extractor, which processes tens of millions of customer interactions monthly. The goal was to maintain high accuracy in identifying deal and account information without sacrificing quality.

• Method: The FDR team built a unique training dataset and applied SFT and DPO techniques over two weeks.
• Result: The specialized model matches Claude Opus 5.5 in speed and accuracy according to Gong’s evaluations, but at a small fraction of the cost.

Deel: Surpassing Frontier Accuracy on Payroll QA

Deel, a global HR and payroll platform, required a model that could provide complete answers to complex, jurisdiction-specific payroll questions where partial answers are considered wrong.

• Method: Using expert-verified data from Emet by Deel, the team applied a two-step recipe: SFT on pruned rollouts followed by SDPO (self distillation policy optimization).
• Result: Answer accuracy reached 72% on an 8-country payroll benchmark, sitting 8 points above Claude Opus 5.5, while costing 33x less per query.

Alice: Enhancing Attack Resilience

Alice, an AI safety company, sought to improve model resilience against indirect prompt injection attacks in environments like Slack and GitHub.

• Method: The team used Alice’s “Rabbit Hole” dataset and RL gyms, applying SFT and DPO followed by GRPO (Online RL) to increase diversity and performance.
• Result: The post-trained model improved attack resilience by 41 points without harming overall task performance, outperforming GPT-6.1 Sol.

Implications for AI Independence

These collaborations highlight a shift toward sovereign specialized intelligence. By post-training open foundation models like Nemotron 3.5 Super VL on proprietary data, enterprises can achieve specific performance goals—such as cost reduction, higher accuracy, or security—while maintaining full control over agent behavior and data. The entire process, from baseline evaluation to specialized deployment, took just three weeks.