Mesa’s reverse-engineered Etnaviv NPU driver now runs the YOLOX object detection model on Vivante NPUs without proprietary vendor blobs.
Announced on 22 August 2026, the reverse-engineered, open-source Etnaviv Linux graphics/NPU driver stack can now successfully run the YOLOX real-time object detection model. The work was brought up by Tomeu Vizoso, a prominent open-source graphics engineer who leads Etnaviv’s NPU/AI driver developments and Mesa’s Teflon (TensorFlow Lite framework) efforts.
The capability is enabled on Vivante VIP NPUs, such as the NPU IP embedded in the NXP i.MX8M Plus system-on-chip (SoC). Support for YOLOX represents a substantial capability jump over the driver’s previously supported flagship model, SSDLite MobileDet, providing significantly higher detection performance and precision.
To execute YOLOX smoothly, additional neural network operations, layer types, and driver features had to be reverse-engineered and implemented directly into the Etnaviv NPU driver/userspace software layers within Mesa.
The enablement serves as a major demonstration of running modern, anchor-free, high-performance object detection pipelines completely on a mainline open-source driver stack on embedded ARM hardware, bypassing proprietary vendor blobs and driver closed-source dependencies.
