Navan brings NVIDIA Nemotron 3.5 Super VL into live business travel AI agents

The travel and expense platform says it fine-tuned and deployed the open model in three weeks, reporting lower latency and stronger results in its own tests.

KEY TAKEAWAYS

• Navan says NVIDIA Nemotron 3.5 Super VL now powers several of its AI travel agents in production.
• The company trained and deployed its fine-tuned version three weeks after receiving early access.
• In Navan’s internal evaluations the model matched or beat alternatives on six of seven measures and had the lowest latency.
• The approach combines open and proprietary models while keeping selected systems self-hosted for operational control.

Navan has put a customised version of NVIDIA’s newly released Nemotron 3.5 Super VL artificial intelligence model into live use across several of its business travel agents, marking another step in the push to make AI-powered travel assistance faster and less costly to operate.

The Palo Alto-based travel and expense platform said on 16 October 2026 that it trained and deployed the model within three weeks of receiving early access from NVIDIA. The agents are now handling customer conversations, although Navan has not disclosed how many interactions they serve, which specific travel tasks have moved to the new model or quantitative changes in operating cost.

Adapting an open model to business travel

Rather than relying on one AI system for every request, Navan says its strategy is to assess models against individual travel workflows, fine-tune selected systems using its travel expertise and run them on infrastructure it controls. The company continues to use both open and proprietary models for agents designed to understand traveller preferences, apply corporate policies and select tools for relevant tasks.

For the latest deployment, Navan fine-tuned Nemotron 3.5 Super VL using the vision-language model LoRA training recipe in NVIDIA NeMo AutoModel. It then compared the adapted model with its existing self-hosted production baseline and other contenders using the same prompts, tools, scenarios and evaluation criteria.

Navan reports that the fine-tuned model delivered its best overall result, matching or outperforming competing models on six of seven measures and recording the lowest latency in its internal tests. It also said the deployment reduced costs. Those are company-reported findings: Navan did not release the underlying scores, full benchmark dataset or a numerical cost comparison in the announcement, so the results cannot be independently assessed from the information provided.

Ilan Twig, Navan’s co-founder and chief technology officer, said reliable AI travel agents require continuous improvement and practical problem-solving beyond simply selecting a base model. He described the early access arrangement as a way to speed development while retaining control over implementation and service quality.

Why the model matters for corporate travel

NVIDIA introduced Nemotron 3.5 Super VL in October as an open, customisable model for agent-based tasks that can involve understanding text, images and documents, following instructions and using software tools. According to NVIDIA’s technical description, the model has 120 billion total parameters in a mixture-of-experts design, with 12 billion active for each token. The architecture aims to balance capability with processing efficiency.

For corporate travel platforms, model responsiveness and consistent policy handling are practical concerns: an assistant may need to interpret a traveller’s request, check company rules and coordinate several systems before producing a useful answer. Navan’s approach illustrates the growing emphasis on adapting AI models to those operational tasks rather than treating a general-purpose chatbot as a complete travel solution.

Navan offers travel booking, expense management and traveller support through its business travel platform. While its new production deployment is confirmed by the company, the announcement does not establish whether travellers will notice measurable changes across all parts of the service.