AI21 customizes NVIDIA models for three partners, citing cost and security gains

AI21 says three weeks of targeted training turned NVIDIA’s Nemotron 3.5 Super VL into specialized models for Gong, Deel and Alice. In its October 15 announcement, the company describes lower-cost sales and payroll systems, plus stronger defenses against malicious instructions—results measured against the particular jobs each partner wanted its model to do.

Training around a business’s definition of “right”

The work starts with an existing model rather than training one from scratch. AI21’s researchers then adapt it for a narrower assignment, using custom training data and evaluations built around customer objectives. The company delivered the projects through its Lab-as-a-Service offering and Harness and Model training platform.

Each partner supplied a different target. Gong wanted cheaper extraction of sales information without losing quality. Deel wanted more complete payroll answers. Alice wanted agents that better resisted attacks while still completing their work. AI21 used supervised fine-tuning and preference-based training across the projects, adding online reinforcement learning for Alice’s security task.

Sales records and payroll answers need different training

Gong’s Data Extractor reviews calls, emails and web conferences to identify deal and account information. It updates fields that can sync to customer relationship management records, including deal stages and health scores. Gong processes tens of millions of interactions monthly, making the cost of each extraction consequential at scale.

Within two weeks, the teams assembled a training dataset, combined supervised fine-tuning with preference optimization, and built evaluations for edits to those records. Higher-risk metrics received extra checks. AI21 says Gong’s evaluations found that the resulting model matched Claude Opus 5.5 on speed and accuracy at a fraction of the cost.

Deel’s problem was completeness rather than extraction. Payroll answers depend on local tax rules, employment contracts and company policies. AI21 says initial testing showed Nemotron already held much of the necessary knowledge, but needed training to deliver the expected detail and style without extra prompting.

The team used expert-verified data and benchmarks from Emet by Deel, its AI evaluation lab. The training followed two steps:

1. Supervised fine-tuning on selected model-generated responses.
2. Self-distillation policy optimization, with a teacher given detailed instructions about the desired behavior to guide the model’s training.

On Emet’s eight-country payroll benchmark, AI21 reports 72% answer accuracy—eight percentage points above Claude Opus 5.5—at one-thirty-third the cost per query. That remains a task-specific result: 72% accuracy is not error-free payroll advice, and the comparison comes from AI21’s account of its partner’s benchmark.

Alice trained for hostile instructions

Alice builds reinforcement-learning environments where agents train against security threats in software platforms. Its Rabbit Hole dataset draws on real-world threat intelligence and contributions from more than 150 staff security researchers and expert-network members. The collaboration targeted indirect prompt-injection attacks in Slack and GitHub: malicious instructions encountered by an agent during its work.

Researchers first evaluated Nemotron in Alice’s training environment. They then used supervised fine-tuning and preference optimization to improve initial performance and response diversity, before adding online reinforcement learning. AI21’s security results show a 41-point improvement in attack resilience without reducing overall task performance. That measures resistance to the tested attacks, not a promise that the model cannot be manipulated.