Architecture supports on-device inference without cloud connectivity

Ceva has partnered with embedUR systems to launch a dedicated instance of the ModelNova platform for Ceva’s NeuPro NPU family. The platform provides AI models, datasets, and use-case blueprints optimized for NeuPro-Nano NPUs and NeuPro-M NPUs for edge AI inference applications.

ModelNova hosts models from both companies for the development of wearables, hearables, smart home devices, and AIoT applications using NeuPro-Nano NPU IP. The platform supports industrial automation, robotics, smart cameras and AR/VR applications on NeuPro-M processors. Development roadmap includes transformer-based architectures, large language models, and generative AI implementations.

The platform integrates models optimized with TensorFlow Lite Micro into Ceva’s NeuPro Studio SDK. Models support object detection, face recognition, pose estimation, audio denoising, keyword spotting, and anomaly detection. Each model includes performance metrics, execution script,s and dataset descriptions to reduce evaluation and integration time.

The architecture enables on-device inference without cloud connectivity for applications requiring low latency, data privacy, and autonomous operation. NeuPro-Nano targets ultra-low-power inference while NeuPro-M handles compute-intensive edge AI workloads. The platform approach reduces development overhead by providing pre-optimized models for NeuPro hardware.

The ModelNova instance for Ceva’s NeuPro NPU family is available with ongoing expansion of models and use-case blueprints.