The Pulse
Navan Puts NVIDIA’s Nemotron 3.5 Model Into Travel Agents
Navan says NVIDIA’s Nemotron 3.5 Super VL now powers several of its AI travel agents in production. The company fine-tuned the model in three weeks and says it beat or tied competing models on six of seven internal measures.
AI.info Team · Written with AI ·
“Building reliable agentic travel systems involves much more than choosing a model, but continuously iterating to improve and solve real-world problems,”
Ilan Twig, co-founder and CTO of Navan
Three weeks from early access to production
Navan says NVIDIA’s Nemotron 3.5 Super VL now powers several of its AI travel agents, putting the model into customer conversations on its business travel and expense platform. The company announced the deployment on October 16, saying its team trained and launched the model three weeks after receiving early access and starting work with NVIDIA.
The move fits Navan’s open-model strategy: test models for particular travel tasks, adapt them using the company’s travel expertise, then host them itself. Navan says self-hosting gives it control over performance, cost and reliability, while its system also combines open and closed models.
Navan’s test favored its fine-tuned model
Navan evaluated Nemotron against its self-hosted production baseline and other models using consistent prompts, tools, travel scenarios and criteria. After fine-tuning with NVIDIA’s NeMo AutoModel VLM LoRA recipe, Nemotron delivered the strongest overall result in the company’s internal evaluation, beating or tying every competing model on six of seven measures and recording the lowest latency, Navan said.
In a technical account published alongside the announcement, Navan identifies Qwen 3.8-27B as its main production baseline for the comparison. The company says Qwen had previously outperformed Gemini 3 and 3.5 Flash in its travel workflows. Navan describes its tests as simulations based on production conversations, measuring tool use, task completion, policy and constraint following, grounding in real information, conversation quality, output integrity and reasoning.
One model serves five travel specialists
Navan says it fine-tuned one model using the combined experience of five agents: an orchestrator for general support, plus specialists for hotels, flights, cars and trains. The agents handle distinct tasks but can share context, so a disrupted flight request, for example, can be passed to a flight specialist without losing information gathered earlier.
The company’s technical account says the final training set contained more than ten times as many examples as its initial discovery set and nearly 2,000 validation tests. It also describes roughly 10 million tokens of supervised training information, compared with about half a million in the discovery set. Navan used LoRA, a method that adjusts a smaller part of a model rather than retraining all its parameters.
The benchmark’s limits matter
Although Nemotron 3.5 Super VL supports text, images, documents, graphical interfaces and video, Navan says it focused this evaluation on text-based travel tasks and industry-specific tools. The company’s account also describes a model with roughly 120 billion parameters that activates about 12 billion for each step, and says multi-token prediction helps reduce the steps needed to generate an answer.
The reported edge is Navan’s own evaluation, not an independently published benchmark: the announcement gives no numerical scores for the seven measures. It also says the model is already powering customer conversations at lower cost, but does not quantify the savings. For travel agents, the test is not simply whether a model can suggest a flight; it is whether the full system can use live tools, follow company rules and complete a change reliably. Navan’s announcement puts Nemotron into that workflow, while leaving the size of its performance and cost gains unreported.