A camera that describes what it sees: local AI on Raspberry Pi

From an image to words

A camera provides an image; a vision-language model can turn it into a description. Raspberry Pi demonstrates that combination in the video above: the AI HAT+ 2 runs a model that interprets the camera input. The scene is processed on the device, without sending it to a cloud AI service. It is a concrete way to understand local AI: giving a small project eyes.[1]

What the board adds

The accessory mounts on a Raspberry Pi 5 and adds a Hailo-10H chip with 8 GB of dedicated model memory. That memory does not expand the computer’s general RAM. The chip handles AI calculations, leaving the Pi’s processor available for other project tasks. In its February explanation, Hailo identifies interpreting an image and producing a short answer as a favourable workload; that does not mean every conversation will run faster.[2][4]

Choose the task before the power figure

This distinction matters when a project must handle a camera alongside other functions. For text generation without images, the board does not always win: in his 15 January test, Jeff Geerling found the Pi 5 CPU faster with most of the models he compared. Those results describe that configuration, not all current software. The useful question is which part of your task keeps the processor busy and whether the accelerator can take it over.[3][4]

What you need to prepare

Connecting the accessory is only the start. The documentation requires the relevant software and models prepared for the Hailo-10H. Camera applications integrate with tools such as Picamera2, but an arbitrary model downloaded from the Internet does not automatically become compatible. For Raspberry Pi project builders, a focused task and a demonstration already working on that hardware provide a practical starting point.[2]

Four models for different tasks

Hailo’s catalogue includes these variants prepared for the accelerator. The two text models can run through Hailo-Ollama; vision and speech models are listed for Hailo’s Python or C++ programming interfaces. The compiled package must match the version of HailoRT, the software connecting the application to the chip.[5]

• Qwen2.5-1.5B-Instruct — Conversation and text responses.
• Qwen2.5-Coder-1.5B-Instruct — Help writing or reviewing code.
• Qwen3-VL-2B-Instruct — Interpret images alongside a written question.
• Whisper-Base — Turn recorded speech into text.

Where to buy it and what it costs

AI HAT+ 2 prices checked on 7 October 2026, including VAT. These are accessory prices: the Raspberry Pi 5 is sold separately. Each shop’s page lets you check availability and the final shipping cost for your address; prices and terms can change.[7][6]

• RaspiPC · €219.84 ↗ — The page shows limited stock and shipping from €2.48.
• Tiendatec · €221.95 ↗ — Advertises immediate dispatch availability and free shipping to mainland Spain.