The Raspberry Pi 5 has another way to push pixels through a neural network. Sixfab has released the AI HAT+, a PCIe expansion board built around the DEEPX DX-M1 accelerator and aimed squarely at local computer vision workloads like object detection and segmentation.
Unlike the M.2 form factor that DEEPX silicon usually ships in, the chip here is soldered directly to the HAT. The board plugs into the Pi 5 over a 16-pin FFC cable carrying a single lane of PCIe Gen 3, and pulls all of its power from the 40-pin GPIO header at 5V / 3A. No auxiliary connector is required, but Sixfab is clear that the official 27W PSU is mandatory, since the older 15W brick will not feed the combined 13 to 15W draw of a Pi 5 plus a fully loaded NPU. Cooling is passive by default, with a JST fan header included on the 65 x 56.5 mm (2.6 x 2.2 inches) HAT+ compliant PCB for anyone running sustained inference.
Two variants are on offer. The flagship uses the DX-M1M with 25 TOPS of INT8 throughput and 1 GB of LPDDR4X dedicated to the NPU, while the cheaper model drops to the DX-M1ML at 13 TOPS with 512 MB of memory. Software setup leans on the Pi 5's HAT+ EEPROM for auto-configuration on Raspberry Pi OS (Trixie), with the rest handled by installing the dxrt-runtime package from Sixfab's APT repository. The runtime exposes both Python and C++ APIs, and the documentation points to a model zoo with prebuilt YOLOv8, MobileNet, and ResNet binaries. Custom networks can be exported to ONNX and compiled to the DXNN format using DEEPX's DX-COM toolchain. The underlying PCIe kernel driver is published as open source on GitHub, and DEEPX also maintains a Yocto layer for anyone integrating the runtime into a custom embedded Linux image. A formal partnership with Ultralytics adds a native format=deepx export target to the Ultralytics toolchain, so deploying a custom YOLO model to the DX-M1 is a single command away.
It is worth being clear about what this board is not. The DX-M1 has no transformer decoder support and only modest on-package memory, so generative AI, LLMs, and VLMs are off the table. That is the territory the Raspberry Pi AI HAT+ 2 with the Hailo-10H accelerator, 40 TOPS, and 8 GB of dedicated memory was designed for. Sixfab says LLM support is on DEEPX's silicon roadmap, but has not committed to a timeline. For object detection and image processing pipelines, the AI HAT+ slots in alongside the original Raspberry Pi AI HAT+ with Hailo-8.
The 13 TOPS DX-M1ML variant sells for $63 (€58) and the 25 TOPS DX-M1M for $90 (€83), both available now from the Sixfab store. The company is also teasing an Edge AI Expansion Board that pairs the same accelerator with NVMe SSD storage and LTE/5G cellular on a single HAT, though full specs and pricing have not been published yet.
