How Navan is building more reliable AI travel agents with NVIDIA Nemotron Open Models

Leveraging self-hosted models to provide greater control over reasoning quality, speed, and efficiency

Most agentic AI systems aim to be smart even at the expense of being slow. Since Navan is a travel company, our agentic system has to be both smart and fast, which creates an exciting engineering puzzle to solve. Agents (whether human or AI) have to be capable enough to understand what a traveler wants, with the ability to search live inventory and use the right tools to make an instantaneous change.

An alert saying a flight was canceled is not enough. Navan’s system needs to find alternative flights, explain the options, address the traveler’s concerns, and help complete the change. The agent needs the human trait of cognitive flexibility, which remains a challenge for many agentic systems. And the models that best handle cognitive flexibility can use tens of thousands of tokens, which can take substantial time.

Our models must be both intelligent and efficient. We require that they solve problems with a fraction of those tokens (several hundred for the entire response). These constraints make the model selection process absolutely critical. Conventional approaches fall short and specialized evaluations and benchmarks need to be created. That’s exactly what the Model Foundry Team at Navan does.

We have been training proprietary models using our extensive dataset of millions of trips and support interactions. This has helped us achieve superior accuracy, enhance performance, and reduce costs. To broaden our model portfolio, we recently gained early access to NVIDIA Nemotron 3.5 Super VL. Within three weeks, we fine-tuned the model and deployed it to production, where it currently drives customer interactions at a reduced cost.

With early access to Nemotron 3.5 Super VL, our team used the tools we have built to test whether Nemotron 3.5 Super VL can outperform our current models. We benchmarked how reliably it performs the core tasks required in real travel and how much domain-specific fine-tuning could improve it. We used Qwen 3.8-27B as our primary comparison because it was already Navan’s self-hosted production baseline. Qwen had previously outperformed Gemini 3/3.5 Flash in our travel workflows, so the practical question was whether Nemotron 3.5 Super VL could deliver comparable or better quality with greater efficiency.

The results were significant. Even without any fine-tuning, Nemotron 3.5 Super VL compared well on intelligence and outperformed Qwen in efficiency during our tests.

On this foundation, fine-tuning improved the model’s intelligence within strict response-length constraints. Using only a few hundred tokens, orders of magnitude fewer than competing high-reasoning models (and Nemotron 3.5 Super VL base). The fine-tuned model was the most intelligent all while delivering the shortest latency.

Building agentic systems vs. chatbots

When a traveler says, “My flight today was canceled. Find me a way to get to London tomorrow,” any travel agent (human or AI) should do more than write a response. To be effective, the agent must understand the request and its consequences, determine which tasks need to be completed, and assign each one to the right AI specialist. These tasks may include:

• Finding alternative flights
• Changing the hotel and car reservations
• Considering the company’s travel policy
• Explaining the options to the traveler
• Confirming each decision with the traveler

It helps to think of Navan’s AI as a team. One agent coordinates the work, while specialized agents handle different parts of the trip:

• Orchestration and general support
• Hotels
• Flights
• Cars
• Trains

The orchestrator may recognize that a canceled flight requires help from the flight agent. The flight agent then searches for alternatives, explains the choices, and helps complete the change. For example, a hotel agent is responsible for hotel tasks, not flights. Each agent has a specific job, but they share enough information to manage the full travel journey.

Taking Nemotron 3.5 Super VL to production

Early access to Nemotron 3.5 Super VL gave us a model with the potential reasoning ability and efficiency needed to support our team of specialized agents. The currently used open-weight model in production, Qwen 3.8-27B, is the main model we tested against. We compared Nemotron 3.5 Super VL to it using the same prompts, tools, travel situations, and scoring system. We measured tool use, task completion, constraint following, grounding in real information, conversation quality, output integrity, and reasoning. So the practical question was whether Nemotron 3.5 Super VL could improve on the model we were already running.

Evaluating model intelligence and efficiency with real-world scenarios

The evaluation process uses simulations of conversations that are based on production scenarios. These measure precisely the key performance indicators of intelligence and efficiency. In our evaluation, the fine-tuned Nemotron 3.5 Super VL model had the strongest overall performance. It beat or tied every competing model on six of seven measures and performed especially well at using tools, completing tasks, staying grounded in real information, and holding a useful conversation.

Nemotron 3.5 Super VL gave us the balance we needed for long-running travel workflows: enough capacity for complex reasoning, while activating only part of the model for each request. Nemotron 3.5 Super VL has access to roughly 120 billion parameters, or model settings, but activates only about 12 billion for each step. Think of it as having a large team but calling on only the specialists needed for each part of a problem.

This gives the system access to the capacity of a much larger model without making every request use the full model. The result is a useful balance of reasoning quality, latency, and serving cost.

Nemotron 3.5 Super VL also uses multi-token prediction. Instead of generating one piece of an answer at a time, it can suggest several pieces and check them together. That can reduce the number of steps needed to produce an answer and make better use of the hardware, without changing the result.

Fine-tuning for real travel

A general AI model knows a lot about language, but it does not automatically know how a travel platform works. It needs to learn how to use travel tools, interpret company policies, handle agent handoffs, and recover when a search fails. Super VL supports text, images, documents, GUIs, and video. For this evaluation, we focused on text-based travel tasks and tools that are tailored to the industry.

We approached fine-tuning in two stages. First, we ran a discovery set using several thousand training examples. This helped us adapt the training process for text-based tools used in travel conversations.

The final set included more than ten times the training examples relative to the discovery set and nearly 2,000 validation tests. More importantly, it contained roughly 25 times more supervised training information, i.e. approximately 10 million tokens compared with approximately half a million. This gave the model many more examples of how to use travel tools, follow policies, and complete travel tasks.

We used LoRA fine-tuning with NVIDIA NeMo AutoModel open library, which allowed us to adapt the model for travel while changing only a small part of its parameters. In simple terms, we taught the model new travel-specific skills without rebuilding it from scratch, which saved significant time and costs.

One model, five agents

Instead of building separate models for hotels, flights, cars, and trains, we fine-tuned one model using the combined experience of all five agents.

This helps the model understand the full travel journey while providing clear boundaries. For example, a flight problem may need to be handed off to a specialist that can use the information already collected. Each agent remains responsible for its own tools and decisions but can share context. This allows useful skills, such as applying policy, collecting missing information, and recovering from failed searches, to benefit the entire system.

Testing the complete system

A model can perform well in a simple test and still fail when connected to real tools. That is why we test the complete travel assistant under the same prompts, tools, scenarios, starting conditions, and scoring rules.

The model must choose the right tool, follow the rules, complete the task, use accurate information, and recover safely when something goes wrong. A faster model or a lower training loss is not enough. It has to work as part of the entire system.

Why self-hosting matters

Self-hosting does not mean we expect one model to solve every problem forever. It means we can test models ourselves and replace them when a better one comes along.

Because Navan controls how the model is deployed, customized, tested and updated, we have greater control over speed, cost, and reliability. We can improve the model without rebuilding the entire system around it.

The future of AI in travel will not be defined by chatbots that give clever answers. It will be defined by systems that understand what a traveler wants, follow the rules, use the right tools, and get the job done.

Nemotron 3.5 Super VL is one building block for that future. And because we control the technology behind it, we can keep improving the experience as the technology changes.