Google AI Edge and Raspberry Pi have published a joint announcement on running on-device AI on the Raspberry Pi 5 with LiteRT, Google’s on-device inference runtime, and the Gemma family of open models. The announcement — published on the Google for Developers blog and the Raspberry Pi website — shows a Raspberry Pi 5 running language, vision, and speech models entirely locally, including the Reachy Mini robot demo, where object detection, speech recognition, and a Gemma 4 E2B language model work together in real time with no cloud dependency.

Igalia contributed the GPU driver side of this work as part of our long-standing collaboration with Raspberry Pi on its open-source graphics stack. José María Casanova, from Igalia’s Graphics Team, enabled the experimental WebGPU support in Mesa’s V3DV Vulkan driver that LiteRT’s WebGPU backend needs to run on the VideoCore VII GPU, through the Dawn WebGPU implementation. This experimental support makes it possible to run inference on the GPU alongside the CPU, so continuous workloads can execute without competing for the CPU cycles that language-model inference and application logic need — as the Reachy Mini demo does, running its object detection on the GPU while Gemma runs on the CPU. It can be enabled with the V3D_WEBGPU_OVERRIDE=1 environment variable in the updated official Raspberry Pi OS Mesa package.

As with all of our graphics work, these improvements were developed in upstream Mesa, so they benefit every Vulkan and WebGPU user on Raspberry Pi devices, not just AI workloads. Work continues on the V3DV driver to improve compute and 16-bit floating-point performance, bringing more of the platform’s capabilities to on-device AI.