FEV and Microsoft bring offline generative AI to vehicles with small language models on NVIDIA
FEV is collaborating with Microsoft to put small language models inside cars on NVIDIA hardware, enabling voice, text, and gesture controls that work without internet.
- SAE Levels 3-5
- 5,600 employees
- more than 45
What Happened
FEV is collaborating with Microsoft to bring generative AI into vehicles using NVIDIA GPU-accelerated compute and AI model microservices. The cooperation aims to enable multimodal voice, text, and gesture interactions directly in the vehicle, independent of a permanent internet connection. The focus is on small language models such as Microsoft's Phi-4-mini-instruct in Microsoft Foundry, powered by NVIDIA DRIVE AGX accelerated compute. The solution lets vehicle functions like the dashboard or individual vehicle profiles be configured by voice command.
Because inference happens on-board, central functions remain available even with limited or no internet connection. Embedded small language models also reduce backend and infrastructure costs, since cloud-based large language models can be supplemented or partially replaced depending on the use case. FEV says this helps OEMs economically scale software-defined vehicle functions.
- Automated and autonomous driving (SAE Levels 3-5): Multimodal GenAI models improve recognition of objects, traffic situations, and driving paths, especially in complex urban environments and edge case
- Driver and passenger monitoring: Embedded GenAI increases performance in detecting fatigue, distraction, or unusual behavior, and adds robustness through local availability, also as a backup to cloud
- Personalized vehicle and HMI configuration: Vehicle functions and user interfaces can be intuitively adapted by voice command for different driver profiles or usage scenarios, without dependence on ex
“Our collaboration with Microsoft and NVIDIA showcases how small, efficient language models can transform in-vehicle experiences, delivering powerful functionality without the overhead of larger systems.”
The underlying architecture is multimodal and processes speech, text, and visual information. To optimize the Phi-4-mini-instruct model, FEV used synthetically generated data curated with NVIDIA NeMo in the fine-tuning process, then integrated and deployed the model on NVIDIA DRIVE AGX. AI functions are operated as modular software services inside the vehicle. FEV has developed a dashboard configurator showcase where a locally deployed small language model updates the dashboard using natural voice commands, reducing reliance on continuous cloud connectivity; completed in just a few weeks, it demonstrates strong real-time performance. FEV plans to gradually supplement or replace cloud-based AI functions with models that run locally.
“By combining advanced AI frameworks with domain- and task-specific optimizations, FEV and Microsoft are shaping the future of intelligent, voice-driven interfaces that meet the high standards of automotive deployment.”
FEV contributes expertise in software-defined vehicles, AI integration, and additional development solutions, with a portfolio spanning from use case definition through architecture and system design to validation and industrialization. Results of the collaboration are currently being tested in demo vehicles and will serve as the basis for further customer projects and near-series applications later this year. FEV is also focused on bidirectional communication between the driver or passenger and the vehicle across a wide range of use cases.
Why this matters
This matters because it lets drivers use voice commands to configure vehicle functions even with no internet, and helps automakers reduce backend costs by supplementing or partially replacing cloud-based AI. FEV says this approach helps OEMs economically scale software-defined vehicle functions.
Terms in This Story
- Small language model (SLM)
- A compact AI model designed to run efficiently on limited hardware, such as inside a vehicle, rather than in massive data centers.
- Large language model (LLM)
- A powerful AI model typically hosted in the cloud that processes and generates natural language but requires substantial computing resources.
- Inference
- The process of running a trained AI model to make predictions or generate responses, such as interpreting a voice command.
- Fine-tuning
- The process of further training a pre-trained AI model on additional data to adapt it to a specific task or domain.
Summarised from the linked release; details can be imperfect — always verify against the original source.