Navan is accelerating its move toward model independence by integrating NVIDIA Nemotron 3.5 Super VL into its production AI travel agents. By leveraging an open-model strategy, the NASDAQ-listed platform aims to maintain granular control over performance and costs. This deployment follows a rapid three-week development cycle after the company received early access to the NVIDIA model.
Navan’s Open-Model Strategy and NVIDIA Integration
The company is positioning its AI travel agents around a hybrid architecture that combines both open and closed models. To optimize for specific travel tasks, Navan utilizes the NVIDIA NeMo AutoModel VLM LoRA training recipe to fine-tune the Nemotron 3.5 Super VL model. This approach allows the company to self-host models, which Navan claims helps maintain control over reliability and operational expenses. According to internal evaluations, the fine-tuned Nemotron 3.5 Super VL outperformed or matched competing models across six of seven measured criteria. Most notably, the model delivered the lowest latency during testing, which Navan is now utilizing to power live customer conversations in a production environment.
Performance Metrics and Operational Efficiency
Navan’s internal testing compared Nemotron 3.5 Super VL against its existing self-hosted production baseline and other available models. The evaluation used consistent prompts, tools, and specific travel scenarios to ensure data integrity. The results suggest that the fine-tuned model provides significant efficiency gains, with Navan stating the deployment allows for customer interactions at a lower cost. By adapting these models with proprietary travel expertise, Navan intends to build agentic systems capable of understanding traveler intent, adhering to corporate policies, and utilizing necessary tools. This deployment highlights a shift toward specialized, fine-tuned models rather than relying solely on general-purpose, closed-source large language models for enterprise-grade automation.
Key Takeaways
- Navan deployed NVIDIA Nemotron 3.5 Super VL into production just three weeks after receiving early access.
- The model achieved the lowest latency and met or exceeded competing models on six of seven internal performance measures.
- Navan uses the NVIDIA NeMo AutoModel VLM LoRA training recipe to fine-tune models for specific travel tasks.
TechInsyte's Take
In our view, Navan’s rapid deployment signals a maturing enterprise approach to AI: moving away from "black box" dependency toward specialized, self-hosted intelligence. By fine-tuning open models like Nemotron 3.5 Super VL, Navan is attempting to solve the dual challenge of latency and cost-efficiency. This strategy suggests that for high-stakes enterprise workflows, the ability to control the underlying model architecture is becoming as critical as the AI capabilities themselves.
Questions & Answers
How does Navan maintain control over its AI performance and costs?
Navan utilizes an open-model strategy where it evaluates specific models for travel tasks, adapts them using proprietary expertise, and self-hosts them to manage reliability and expenses.
What specific technical method did Navan use to fine-tune the NVIDIA model?
Navan applied the NVIDIA NeMo AutoModel VLM LoRA training recipe to fine-tune the Nemotron 3.5 Super VL model for its specific use cases.
What were the results of Navan's internal model evaluation?
The Nemotron 3.5 Super VL delivered the strongest overall performance, beating or tying competing models on six of seven measures, including achieving the lowest latency.
How does Navan's AI architecture handle complex travel requirements?
Navan combines open and closed models to create agents that can understand traveler needs, follow corporate policies, and utilize the appropriate tools for task completion.
Source: Navan
