AI21 Post-Trains Nvidia Nemotron 3.5 Super VL for Gong, Deel and Alice

Owning specialized intelligence instead of renting it

AI21 frames the project as part of a shift in which companies want to own specialized intelligence rather than rent general models. The company says its Lab-as-a-Service offering and its proprietary Harness and Model training platform let firms make that transition "in weeks, not quarters."

The base model is NVIDIA Nemotron 3.5 Super VL, described by AI21 as an open multimodal model for long-running agents that need to see, reason and act. It is designed to be post-trained on an organization's own data, tools and policies. AI21 says motivations for taking this route differ: for some companies it is cost, for others control over data or a differentiating edge.

Gong and Deel: cost and accuracy results

Gong. The revenue AI company built an agent called Gong Data Extractor, which reviews customer interactions such as calls, emails and web conferences and stores key deal and account information as fields that can sync to CRM records. Gong processes tens of millions of interactions a month, so cost compounds quickly, according to AI21. The goal was to cut cost without hurting quality.

AI21 says the teams built a training dataset, used a mixture of SFT and DPO training, and set up evaluation infrastructure to gate quality on CRM edits, with extra layers for high-risk metrics. This took just two weeks. Gong's own evaluations found that the specialized model matches SOTA models such as Claude Opus 5.5 in speed and accuracy, at a small fraction of the cost, AI21 reports.

Deel. The global HR and payroll platform, which supports more than 40,000 companies across 150+ countries, wanted an open model that gives complete payroll answers unprompted. AI21 notes that frontier models often leave part of an answer out, and in payroll a partial answer is a wrong answer.

The team used expert-verified training data and benchmarks from Emet by Deel. The recipe had two steps: SFT on pruned generated rollouts, then SDPO (self distillation policy optimization), in which the teacher gained hindsight by conditioning on detailed instructions about the desired behavior. According to AI21, answer accuracy rose to 72% on Emet's 8-country payroll benchmark, 8 points above Claude Opus 5.5, at 33x lower cost per query.

Alice: improving attack resilience

Alice, an AI safety and security company, builds RL gyms where customers can train production agents to counter security threats on the software platforms they use. Its "Rabbit Hole" dataset draws on real-world intelligence about threat actor behavior and input from 150+ staff security researchers and expert network members. The post-trained model is aimed at detecting indirect prompt injection attacks in both Slack and Github.

AI21 began by evaluating Nemotron 3.5 Super on Alice's RL gym, then applied SFT and DPO to improve initial results and increase answer diversity. Once performance and diversity were good enough, the team added GRPO (Online RL) training. AI21 says the result was a 41-point improvement in attack resilience without harming overall task performance, outperforming frontier models such as GPT-6.1 Sol.

What it means for enterprises

AI21 points out that the three tasks span summarization, knowledge-intensive retrieval and attack resilience, yet share a repetitive load of up to millions of queries a month and strict standards for what a correct output looks like. Frontier models cannot always capture those standards, the company says.

In AI21's account, post-training Nemotron 3.5 Super on its platform, directed by its researchers, moved all three tasks from baseline to specialized intelligence in three weeks. It says this reduced cost, created an edge competitors cannot rent and gave full control over agent behavior. The results are reported by AI21 and its partners, and the post does not provide independent verification.