Efficient AI models for automotive applications are at the heart of the collaboration between FEV and Microsoft. Together, the companies are developing high-performance AI capabilities for software-defined vehicles. These are based on small language models, Microsoft technologies and the NVIDIA DRIVE AGX GPU platform. This enables voice, text and gesture interactions to be carried out directly in the vehicle. At the same time, the system remains fully functional even without a constant internet connection.
AI models for automotive applications enable greater independence
The collaboration relies on small language models such as Phi-4-mini-instruct from Microsoft Foundry. The models run locally on NVIDIA DRIVE AGX. This enables drivers to control vehicle functions using voice commands. Individual vehicle profiles and the dashboard can also be configured intuitively. Furthermore, the local models serve as a fallback for cloud-based large language models. This increases reliability and reduces dependence on a constant cloud connection.
Software-defined vehicles benefit from local AI
Local small language models enhance the intelligence of modern vehicles. At the same time, they improve the response times of key functions. As inference takes place directly in the vehicle, core applications remain available even without an internet connection. Furthermore, infrastructure and back-end costs are reduced. This enables OEMs to scale software-defined vehicle functions more cost-effectively.
Thomas Hülshorst, Group Vice President of Intelligent Mobility and Software at FEV, emphasises that small language models enable powerful functions with significantly lower resource requirements. Boris Scholl, Vice President of Engineering at Microsoft, also sees the collaboration as an important step towards intelligent and voice-controlled vehicle interfaces.
Key areas of application for embedded GenAI
As part of this collaboration, FEV is investigating several areas of application with high potential for series production:
- Automated and autonomous driving (SAE Levels 3–5): Multimodal GenAI models improve the detection of objects, traffic situations and driving routes, particularly in complex urban environments and in edge cases.
- Driver and passenger monitoring: Embedded GenAI improves the detection of fatigue, distraction or unusual behaviour and, through local availability, increases the robustness of safety-critical functions – including as a backup for cloud-based systems.
- Personalised vehicle and HMI configuration: Vehicle functions and user interfaces can be intuitively customised via voice commands, for example to suit different driver profiles or usage scenarios – without relying on external cloud infrastructure.
Multimodal system architecture on NVIDIA DRIVE AGX
The architecture processes speech, text and visual information simultaneously. For fine-tuning, FEV uses synthetically generated data with NVIDIA NeMo. The optimised Phi-4-mini-instruct model is then integrated into NVIDIA DRIVE AGX. The AI functions run as modular software services directly in the vehicle. This enables users to benefit from high performance and fast response times.
Dashboard configurator demonstrates the potential of local AI
As a technology demonstrator, FEV developed a voice-controlled dashboard configurator. The small language model, which runs locally, customises the dashboard in real time. This reduces reliance on cloud services. At the same time, response times are significantly improved. The demonstrator was developed in just a few weeks and has already impressed with its strong real-time performance.
In future, FEV plans to gradually supplement or partially replace cloud-based AI functions with locally running models.
FEV is driving forward GenAI for software-defined vehicles
FEV has extensive experience in the fields of software-defined vehicles, AI integration and systems development. The company supports projects from the requirements definition stage right through to industrialisation.
The results of this collaboration are already being tested in demonstration vehicles. They form the basis for future customer projects and production applications. At the same time, FEV is developing intelligent solutions for natural communication between the driver, passengers and the vehicle.
FEV
FEV is a globally leading development service provider in the automotive sector and a driver of innovation across various industries. Professor Franz Pischinger laid the foundations for this by combining his academic and technical background with his vision for continuous progress. Since 1978, the company has been developing technological and strategic solutions for the world’s largest car manufacturers, as well as for customers across the entire transport and mobility ecosystem.
That is why FEV also applies its technological and strategic expertise to other sectors, transferring its forward-looking approach to the aerospace and energy sectors. With its 5,600 highly qualified employees at more than 45 locations worldwide, FEV develops solutions designed to meet not only today’s needs, but also those of tomorrow.
