features, price, technical sheet, photos

Raspberry Pi has announced the launch of its new AI HAT+ 2 accessory, a complementary board for the Raspberry Pi 5 which promises to expand its capabilities for uses that involve generative artificial intelligence systems for a modest price. The board incorporates the Hailo-10H neural accelerator and 8 GB of dedicated RAM, features that according to the company will allow large language models (LLMs) to be run directly on the device without the need for an internet connection. Below these lines we tell you all the details.

Raspberry AI HAT+ 2, technical sheet

A plate for local models

The AI ​​HAT+ 2 represents a leap compared to its predecessor, launched just over a year ago. While the original AI HAT+ focused exclusively on computer vision tasks, such as object detection, pose estimation or scene segmentation, this new model seeks to cover an increasingly growing interest in run language models locally and without needing to be connected to the internet.

According to Raspberry Pi, the board offers 40 TOPS of inference performance (at INT4) thanks to the Hailo-10H chip. The incorporation of 8 GB of LPDDR4X RAM on the board itself is the most notable novelty, as it frees up the Raspberry Pi 5’s processor and RAM for other tasks while the neural accelerator takes care of AI processing.

Several language models with between 1 and 1.5 billion parameters will be available at launch, including DeepSeek-R1-Distill, Llama 3.2, and different versions of Qwen. The company assures that they will soon add larger models and that users will be able to adjust these models to specific use cases using techniques such as LoRA (Low-Rank Adaptation).

Beyond paper

Along with the benefits that the complementary board promises, it is good to know that the Hailo 10H chip that it incorporates is limited to a power of 3W, which means that, although it translates into better energy efficiency if we compare it with the chip that includes the Raspberry Pi 5 itself, this power limitation makes it The add-on board offers lower raw performance than the Raspberry Pi 5 itself.

Additionally, while the extra 8GB of RAM is tremendously useful, it’s worth noting that the 16GB version of the Raspberry Pi 5 board itself also outperforms the add-on board in this regard. We must keep in mind that many medium-sized and quantized language models usually use more than 10 GB of RAM, so the AI ​​HAT+ 2 would be limited to smaller models.

However, this is a very interesting solution to have the Raspberry Pi 5 use the performance of the add-on board to run models while using the resources of the main board for any other task.

Who is it intended for?

The AI ​​HAT+ 2 maintains the computer vision capabilities of its predecessor and, according to the Raspberry Pi, can simultaneously run vision and language models. The board could find its niche in power-constrained applications that need to combine vision processing and language model inference, such as industrial automation systems or IoT devices.

It may also be useful as a development kit for companies planning to integrate the Hailo-10H chip into their own products.

For users simply looking to run language models locally on a Raspberry Pi, the most practical option is still to purchase a Raspberry Pi 5 with 16GB of RAM.

Price and availability

The Raspberry Pi AI HAT+ 2 now It is available for $130.although to buy it in Spain it will depend on the price of the official distributors. Furthermore, in the Hailo GitHub repository and in his Developer Zone You can find examples, tutorials and documentation to get the most out of the board.

Images | Raspberry Pi

In Xataka | Someone has used a Raspberry Pi worth less than 10 euros for a great project: creating their own original Macintosh